Patent Application
20140351182
Some smart devices may learn to perform actions based on received triggers. However, in some circumstances, the learning behaviors of smart devices may not be sufficient or desirable to users. For example, when a new smart device is brought into an environment, a user may want the new smart device to have an accelerated learning rate in order to start its usefulness quickly. Users may need a convenient way to control or adjust the learning capabilities of smart devices.
The various embodiments provide systems, computing devices, non-transitory processor-readable storage media, and methods for modifying learning capabilities within a decentralized system of learning devices. An embodiment method that may be performed by a processor of a computing device that is a learning device may include receiving a signal from a nearby device, determining whether the received signal is a learning modifier signal based on data within the received signal, and modifying one or more learning capabilities of the learning device in response to determining that the received signal is the learning modifier signal. In some embodiments, the method may further include determining whether subsequent learning modifier signals are received, and resetting the modified one or more learning capabilities of the learning device in response to determining that subsequent learning modifier signals are not received. In some embodiments, modifying one or more learning capabilities of the learning device may include enabling a learning mode of the learning device, and resetting the modified one or more learning capabilities of the learning device may include disabling the learning mode of the learning device.
In some embodiments, modifying the one or more learning capabilities of the learning device may include disabling a learning mode of the learning device, and resetting the modified one or more learning capabilities of the learning device may include enabling the learning mode of the learning device. In some embodiments, modifying the one or more learning capabilities of the learning device may include adjusting values of variables used to calculate trigger weights of reflexes based on the learning modifier signal, and resetting the modified one or more learning capabilities of the learning device may include adjusting the values of the variables used to calculate the trigger weights of the reflexes to default values.
In some embodiments, the method may further include initializing a timer, and activating or resetting the timer in response to determining that the received signal is the learning modifier signal or determining that subsequent learning modifier signals are received. In such embodiments, determining whether subsequent learning modifier signals are received may include determining whether subsequent learning modifier signals are received before the timer has expired. In some embodiments, the timer may be set based on data from the learning modifier signal. In some embodiments, the learning modifier signal may include one or more of: a learning rate modifier value that indicates whether the learning device should increase or decrease a rate of learning; a device type that indicates a type of learning device affected by the learning modifier signal; a learning mode active setting that indicates whether the learning device should enable or disable a learning mode; a learning rate modifier type that indicates particular types of calculations affected by learning modifier signals; and a transmission frequency that indicates how often the nearby device will send the learning modifier signals.
Various embodiments may include a computing device configured with processor-executable instructions to perform operations of the methods described above. Various embodiments may include a computing device having means for performing functions of the operations of the methods described above. Various embodiments may include non-transitory processor-readable storage media on which are stored processor-executable instructions configured to cause a processor of a computing device to perform operations of the methods described above. Various embodiments may include a system that may include one or more learning devices configured to perform operations of the methods described above.
The accompanying drawings, which are incorporated herein and constitute part of this specification, illustrate exemplary embodiments of the invention, and together with the general description given above and the detailed description given below, serve to explain the features of the invention.
Various embodiments will be described in detail with reference to the accompanying drawings. Wherever possible, the same reference numbers will be used throughout the drawings to refer to the same or like parts. References made to particular examples and implementations are for illustrative purposes, and are not intended to limit the scope of the invention or the claims.
The word “exemplary” is used herein to mean “serving as an example, instance, or illustration.” Any implementation described herein as “exemplary” is not necessarily to be construed as preferred or advantageous over other implementations.
The terms “learning device(s),” “smart device(s),” and “smart box(es)” are used herein to refer to computing devices capable of learning behaviors from observed information by correlating predefined actions with information related to triggers, such as data indicating user inputs, detected changes in states, received signals or transmissions, and other information that may be obtained at the devices. Learning devices may be configured to store new relationships or correlations between triggers and predefined actions over time. In response to detecting a trigger already correlated to a predefined action, a learning device may perform the predefined action or alternatively perform operations to cause other associated devices to perform correlated actions. Throughout the disclosure, the modifier “smart” may be used to indicate an appliance (e.g., a lamp) is a learning device. For example, the term “smart lamp” refers to a lamp that is configured to be a learning device, is coupled to and controlled by a learning device, or otherwise includes components of a learning device.
The term “event” is used herein to refer to data (e.g., an object or other data structure) that represents an action, condition, and/or situation that has been detected or generated by one or more learning devices. Events may be generated (or otherwise obtained) and stored locally on learning devices in response to obtaining information (referred to herein as “occurrence data”) indicating the occurrence of an action or condition. Occurrence data may include various data describing an action or condition, as well as identifying the device that performed or detected the action or condition, such as device identifiers, timestamps, priority information, gain information, state identifiers, etc. Occurrence data may be received or obtained by a learning device from signals or other information from devices connected to the learning devices (e.g., a sensor directly coupled to a processor or core of a learning device, etc.) or otherwise controlled by the learning device (e.g., a non-programmable lamp, etc.). Occurrence data may also be received or obtained by learning devices from broadcast messages (referred to herein as “event report messages”) received from other nearby devices. For example, after generating a first event based on locally encountered sensor data, a first learning device may broadcast an event report message with occurrence data indicating that the first event has occurred so that a second learning device may be enabled to also generate the first event based on the data within the event report message.
The term “reflex” is used herein to refer to stored information within a learning device that indicates at least a correlation or relationship between a trigger and an action the learning device is configured to perform. The stored information of a reflex may include patterns that may be matched to events generated within a predetermined time window to cause the learning device to perform the action of the reflex and/or adjust persistent data stored in relation to the reflex (e.g., trigger weights). Events may be considered the building blocks of the patterns within a reflex. For example, a trigger pattern stored within a reflex may be comprised of one or more events.
Learning devices may be configured to experience continued configuration through learning processes that are continuous and maintained during normal operations (i.e., configuration is not limited to defined training processes that are separate from normal operations). Such learning processes may emulate biological systems to enable learning devices to be easily configured by observing interactions of a user and/or through intuitive training methods. Learning devices may be easily configured to react in a desired manner in response to events, which may be generated as a result of user actions, changes in state of other learning devices, etc. Through simple repetition, a variety of behaviors can be learned by and thus implemented in a decentralized system of a plurality of learning devices without the need for preconditioning or a programmer interface. Using repeated rewarding training inputs, a user may easily train the learning device to automatically perform predefined tasks in response to various triggers. In a similar manner, a user may easily train the learning devices to stop automatically performing a particular task in response to other triggers by using repeated correcting inputs. Learning devices need not have any in-depth understanding of events or contexts related to events in order to perform actions. Instead, learning devices may simply be trained to match patterns of events sent by one or more devices in order to detect triggers and related, predefined behaviors. Such training is beneficial as it avoids complicated or tedious setup or programming.
Learning devices within the decentralized system may be configured to continually learn new associations between triggers and the execution of various functionalities of the devices. For example, a smart stereo may continually monitor a receiver circuit for received signals from nearby devices that may be paired with the stereo's predefined actions for turning ‘on’, turning ‘off’, changing a radio station, adjusting a volume, etc. However, not all learning may be desired by users. For example, a smart TV may be configured to learn to associate turning ‘on’ (or activating) with various triggers reported by signals from nearby devices. Over time, the smart TV may learn to turn on in response to receiving signals indicating a user has sat on a couch equipped with a sensor (e.g., signals from a sensor that detects pressure when the user sits, etc.), the user has walked into the room (e.g., signals from a motion sensor, etc.), and the user saying “TV ON” (e.g., signals from a device with an attached microphone, etc.). However, the user may prefer to do other activities while sitting on the couch or when entering the room, and so may not want or need the smart TV to activate except in response to him/her saying “TV ON”. As the smart TV may be constantly trying to learn new and useful associations, it may be difficult for the user to control what the smart TV does and does not learn. For these reasons, users may desire to have a means for adjusting the rate that the learning devices actually learn, as well as selectively enabling or disabling learning.
Various embodiments provide devices, methods, protocols, systems, and non-transitory processor-readable storage media for modifying learning capabilities within a decentralized system of learning devices. A special-purpose device, referred to as a “learning modifier device,” may cause nearby learning devices to adjust the rate at which they learn t, such as by inhibiting learning or adjusting parameters that determine whether and how much learning devices strengthen or weaken associations between triggers and actions performed by the learning devices in response to the triggers. When the learning modifier device is within the proximity of learning devices (e.g., smart TV, smart lamp, etc.), signals sent by the learning modifier device may temporarily reconfigure the learning devices to utilize modified operations associated with the learning of new behaviors. For example, when near the learning modifier device, a smart lamp may respond to signals sent by the learning modifier device by requiring fewer ‘on’ event signals to be received from a smart wall switch before learning to turn on. As another example, the smart lamp may require fewer correction signals to learn to stop turning on in response to a smart TV ‘on’ event signal. As another example, the smart lamp may stop or start creating new associations between events indicated in received signals and actions the smart lamp may perform (e.g., turning on). Thus, when the learning modifier device is within range of its signals, learning devices may more quickly or slowly learn to perform (or to stop performing) actions in response to pre-associated triggers and/or simply enable or disable the ability to learn anything.
The learning modifier device may be configured to periodically send messages (referred to herein as “learning modifier signals”) when activated that cause receiving learning devices to adjust their learning capabilities. Such learning modifier signals may be transmitted via wired or wireless media, such as Bluetooth LE broadcast packets or WiFi. In some embodiments, learning modifier signals may include data that indicates the extent to which recipient learning devices should adjust their learning (or learning rate), such as scalars, multipliers, boosters, and/or dampeners that may be utilized by the learning devices when calculating trigger weights (or trigger weight changes) for reflexes. For example, a learning modifier signal may include an integer or floating point value that may be multiplied by, divided into, added to, or subtracted from one or more variables (or values) used by a learning device to calculate a trigger weight calculation, causing the trigger weight to be calculated in a different way than normal for that learning device (e.g., trigger weights increase more each calculation, etc.). In various embodiments, learning modifier signals may cause temporary changes to gains used in calculating trigger weights of reflexes such that a learning device may need to detect (or receive) a higher or lower number of a trigger events before learning to actually perform an associated action.
As an illustration, a smart stereo has learned to associate activating an FM tuner in the smart stereo to both an ‘on’ event signal from a smart wall switch and a pressure event signal from a sensor within a couch. In other words, a first association may be stored as a first reflex with a first trigger weight above a first threshold, and the second association may be stored as a second reflex with a second trigger weight above a second threshold. However, a user may only desire to have the smart stereo activate its FM tuner in response to the ‘on’ event signal from the smart wall switch. In order for this to happen, the second trigger weight of the second reflex may need to be lowered below the second threshold such that the FM tuner does not activate in response to the pressure event signal. Normally, the user may have to repeat a correction procedure several times to “unteach” the association between the pressure event signal and activating the FM tuner. For example, the user may have to repeatedly sit on the couch, wait for the stereo to automatically activate the FM tuner, and then press a ‘correct’ button on the smart stereo until the smart stereo has lowered the second trigger weight for the second reflex below the second threshold. This normal process may be a tedious, time consuming effort that the user may not wish to execute.
Continuing the illustration, the user may use a learning modifier device to adjust the learning rate of the smart stereo and thus speed up the time required to unteach the association between the pressure event signal and activating the FM tuner. In particular, the user may bring the learning modifier device within proximity of the smart stereo, causing the smart stereo to adjust variables used for calculating trigger weights. The user may then sit on the couch, causing the pressure event signal to be received at the smart stereo that triggers the FM tuner activation. The user may then press the ‘correct’ button, causing the smart stereo to calculate a new trigger weight for the second reflex using the variables adjusted based on the learning modifier signals. Thereafter, when the user sits on the couch, the smart device may not activate the FM tuner as the single press of the ‘correct’ button was adequate for unteaching the association due to the adjusted learning rate.
In various embodiments, the learning modifier device may be a smartphone, a dedicated signaling device or transmitter (e.g., a “magic wand” device), or another programmable device capable of broadcasting a signal configured for receipt by nearby learning devices. In some embodiments, the learning modifier device may transmit learning modifier signals in response to receiving other signals or user inputs. For example, the learning modifier device may start or stop broadcasting signals in response to receiving a command over the Internet or a local area network via a WiFi router. As another example, the learning modifier device may start or stop transmitting learning modifier signals in response to a user pressing a start button or loading, accessing, or otherwise activating an application on the learning modifier device. In some embodiments, the learning modifier device may transmit learning modifier signals when connected to a pre-associated user device via short-range signaling, such as via a link with a paired Bluetooth device (e.g., a phone).
In some embodiments, the learning modifier signals transmitted by the learning modifier device may indicate that recipient learning devices should enable or disable (i.e., enter or exit) a learning mode during which the learning devices are capable of generating new associations between triggers and actions and/or adjusting trigger weights of pre-existing reflexes. For example, when a smart stereo is by default not capable of generating new reflexes of associations between triggers and actions to be performed, a user may use the learning modifier device to temporarily place the smart stereo in an enabled learning mode so that the stereo may learn to turn ‘on’ in response to receiving an ‘on’ event signal from a nearby smart wall switch. As a further example, when the user does not want the smart stereo to learn any other associations (and therefore store reflexes), the user may remove the learning modifier device from the room (or turn it off), placing the smart stereo in a disabled learning mode. In this way, when the learning modifier device is removed from a location, the learning devices in the location may be secured from unintended learning. For example, when guests visit, a homeowner may remove the learning modifier device from a TV room in order to stop a smart TV from accidentally learning associations based on the guests' activities.
In various embodiments, some learning devices may or may not be affected by the learning modifier device based on data within the learning modifier signals and/or the proximity of the learning modifier device. For example, when learning modifier signals include data indicating only smart TVs are to adjust their learning rates, a smart wall switch may not adjust its learning rate in response to receiving the learning modifier signals.
In some embodiments, learning modifier signals may be encrypted or otherwise secured to prohibit spoofing or other unintended learning by nearby learning devices. For example, broadcast learning modifier signals may include a hash generated using a secret decoding key that is only shared with a user's smart devices. Such security measures may be beneficial when learning modifier devices are used in multi-family buildings, such as apartments or townhouses.
In the following descriptions, learning devices may be referred to as a smart box or smart boxes, which are particular embodiments of learning devices having the components described below with reference to
As an example, a system 100 enabled by various embodiments may include a wall switch 102 connected to a smart box 103a that transmit signals which enable the wall switch 102 to control responses by other devices (e.g., to turn on a floor lamp 104). The wall switch 102 may be connected to the smart box 103a by a wired connection 110, or the smart box 103a and the wall switch 102 may be combined into a single unit. When the wall switch 102 is toggled, its associated smart box 103a may detect this change in state and emit an event report message via an RF transmission 112, which may be received by any of the other smart boxes 103b-103e within a radius of the transmitting smart box 103a. One such receiving smart box 103b may be connected to the floor lamp 104 via wired connection 110b. By way of example, the floor lamp smart box 103b may be trained to respond to an event report message corresponding to the wall switch 102 being moved to the ‘on’ position by generating an event that causes the floor lamp 104 to be turned on. When the floor lamp 104 is turned on, its smart box 103b may broadcast event report messages that include occurrence data indicating the event and that may be received by other nearby smart boxes 103c-103e as well as the smart box 103a connected to the wall switch 102. Alternatively or in addition, the smart boxes 103a, 103c-103e may include a light sensor that may sense the light from the floor lamp 104 so that turning on of the lamp may be treated as a signal indicating an occurrence/condition/action.
As illustrated in
Although not shown in
A smart box may typically be configured to broadcast or otherwise transmit event report messages indicating events at the smart box, such as actions performed at or by the smart box and/or conditions detected at the smart box (e.g., sensor data). For example, a smart box or a transmitter (a “reporter”) wirelessly connected to the smart box may broadcast a signal including data that indicates that a garage door has been opened. It should appreciated that a smart box may not typically be configured to directly engage with other smart boxes in a location, but instead may merely report occurrence data without soliciting responses and/or without consideration of the operations of other devices. However, in some embodiments, smart boxes may directly communicate with each other via such transmissions 112. For example, a new smart box placed within a location (e.g., a home, office, etc.) may transmit signals to other learning devices within the location to ask for data indicating their favorite (or most frequently encountered) events, and in response to receiving response signals from the other devices, the new smart box may be configured to set a bias.
The system 100 may further include a learning modifier device 150 that is capable of transmitting wired or wireless signals for receipt by the various smart boxes 103a-e within the system 100. In particular, the learning modifier device 150 may send learning modifier signals to the smart boxes 103a-103e via wireless transmissions 112′, such as Bluetooth LE, WiFi Direct, RF, etc. For example, the learning modifier device 150 may be configured to periodically broadcast learning modifier signals that cause the first smart box 103a to increase a learning rate used in calculating trigger weights for reflexes. As another example, the learning modifier device 150 may transmit learning modifier signals that cause the first smart box 103a to disable or enable a learning mode (e.g., become able or unable to generate new reflexes and/or change trigger weights for existing reflexes, etc.).
In some embodiments, the learning modifier device 150 may optionally include a network interface configured to provide a connection to the Internet 152 via the wireless or wired link 153. For example, the network interface may be a transceiver (e.g., WiFi radio, cellular network radio, etc.) capable of communicating with a wide area network (WAN). Based on the connection to the Internet 152, the learning modifier device 150 may communicate with various devices and data sources via Internet protocols, such as a remote server 154 connected to the Internet 152 via a connection 155. The remote server 154 may be a web server, cloud computing server, or other server device associated with a third-party. For example, the remote server 154 may be a server associated with a web portal or a data source that distributes programs, apps (or applications), commands, instructions, scripts, routines, configurations, or other information that may be downloaded and used on the learning modifier device 150.
In various embodiments, the smart box 103b associated with the floor lamp 104 may be trained to energize or cause the floor lamp 104 to turn on in response to receiving a toggle signal (i.e., an event report message including occurrence data indicating the toggle action) from the wall switch smart box 103a by the user manually turning on the floor lamp 104 just before or soon after toggling the wall switch 102 (e.g., within 5-10 seconds). To accomplish such learning, the smart box 103b may recognize when the events related to the wall switch toggle (as reported in the event report message) and the activation of the floor lamp 104 (as reported via occurrence data obtained from the floor lamp 104) occur within a predetermined window of time. This may be accomplished at least in part by buffering events generated from obtained occurrence data for the predetermined window of time, processing and correlating events stored in the buffer, and deleting events from the buffer after that time. For example, the smart box 103b connected to the floor lamp 104 may associate the ‘on’ event of the wall switch 102 with the ‘on’ event of the lamp switch 126 of the floor lamp 104 when the two events are generated or occur within the predetermined window of time, in effect learning that future wall switch 102 ‘on’ events should trigger the activation of the floor lamp 104. In some embodiments, the order of events may be significant, while in some embodiments, the order of events may not matter, and so the order of events may be reversed so long as the events occur (or are generated) within the predetermined window of time. For example, the smart box 103a connected to the wall switch 102 may associate the ‘on’ event of lamp switch 126 of the floor lamp 104 with the subsequent ‘on’ event of the wall switch 102 (e.g., a touch to the touch sensor 119), in effect still learning that future wall switch 102 ‘on’ events should trigger the activation of the floor lamp 104. As described in more detail below, such training may require some repetition to avoid inadvertent learning of undesired behaviors.
As illustrated in
Additionally, the smart box 103 may include a motor driver 140 to perform physical actions on a connected device as a learned reflex action in response to a correlated trigger. For example, if the smart box 103 is connected to a floor lamp and determines based on an event generated in response to a received event report message that the floor lamp should turn on, the processor 132 of the smart box 103 may signal the motor driver 140 to actuate a power switch on the floor lamp. Instead of (or in addition to) a motor driver 140, the smart box 103 may include a relay configured to connect an appliance to an external power supply (e.g., 120 V AC power) as a learned reflex action in response to a correlated trigger.
In some embodiments, the smart box 103 may include a battery 143 (e.g., a rechargeable lithium-ion battery, etc.) coupled to components of the smart box 103. In some embodiments, the smart box 103 may additionally include a wire or other interface 144 (e.g., plugs or prongs for connecting to an alternating current (AC) power outlet, etc.) for receiving electrical current for charging the rechargeable battery 143 or otherwise providing power to the various components of the smart box 103.
The smart box 103 may also receive occurrence data within signals (e.g., event report messages) from another smart box via a signal receiver 142. Data from signals received by the signal receiver 142 may be transported as events to other device components via the event bus 214, such as to the event recorder 206.
A smart box 103 may also recognize an event from the sensor encoder 134, which may communicate the event to other components via the event bus 214. For example, if a user manually turns on a floor lamp connected to the smart box 103, occurrence data indicating that change in state (e.g., turning the light from ‘off’ to ‘on’) may be digitally encoded by the sensor encoder 134 converting the change in state to an event.
A signal transmitter 136 may subsequently transmit occurrence data based on an event received via the event bus 214 so that the occurrence data may also be received by another smart box via event report messages. This may allow the transfer of information about events from one smart box 103 to another, allowing smart boxes to learn from each other and create complex system behaviors based upon behaviors learned by each respective smart box. The retransmission or broadcasting of data related to events (i.e., occurrence data in event report messages) may allow the smart boxes to be daisy-chained together extending the signal range of a given smart box.
The event recorder 206 may receive an event from the event bus 214 and save the event in event pattern storage 204. In some embodiments, the event recorder 206 may receive occurrence data and create an event based on the received data for storage in the event pattern storage 204. An event selector 210 may receive one or more events from the event recorder 206. In response to receiving a particular combination of events, the selector 210 may generate a store pattern command and send the store pattern command to the event recorder 206 instructing it to store the combination of events as a pattern in the event pattern storage 204. In some embodiments, the event selector 210 may receive events directly from the event bus 214.
The operations and interactions of the components with a smart box 103 are illustrated in the following example. A smart box 103 connected to a floor lamp may receive occurrence data indicating a change in state via an event report message from a wall switch, received at the smart box 103 through the signal receiver 142. The smart box 103 via the signal receiver 142 may communicate an event related to the wall switch change in state via the event bus 214 to the event recorder 206. Shortly thereafter, a user may manually turn on the light 124 of the floor lamp connected to the smart box 103, and in response the sensor encoder 134 may convert this change in state to an event and communicate the event via the event bus 214 to the event recorder 206. The event recorder 206 may send the events to the selector 210 as they are received. The selector 210 may process the pattern of events, generated based on the wall switch toggle and the floor lamp's manual on-light occurrence data, with a learning algorithm. After processing the events, the selector 210 may instruct the event recorder 206 to store the pattern of events in the event pattern storage 204 through a store pattern command. The event pattern storage 204 may store the learned association between events as a reflex with a particular weight association. In some embodiments, the event pattern storage 204 may store predetermined patterns and/or events as well, such as patterns or events used to generate correction patterns, reward patterns, trigger patterns, and action patterns.
Depending on the associations between observed events and actions, the selector 210 may work with a gain adjuster 212 to change the weight of an event (e.g., increase the trigger weight of the trigger event) associated with an observed action pattern (e.g., an observation that the user has turned on the floor lamp) and/or other properties related to the equations and/or calculations of weights (i.e., bias, scale, etc.) as described below.
Optionally, the sensor encoder 252 may provide additional events based on the commencement of an instructed action. These additional events may be a confirmation that an instructed event actually occurred (e.g., a light actually came on in response to an ‘on’ action being performed, etc.) and may be processed as reward events (or patterns) to help the smart box 103 learn associations between events and actions.
For example, a data structure 300 for occurrence data may include a format component 301 of “V2.1”, an identification component 302 of “WALLSWITCH102,” and a state component 303 of “ON.” This may represent a data format version of 2.1 on a smart box connected to the wall switch and may represent that the wall switch was toggled from ‘off’ to ‘on.’ Continuing this example, the occurrence data and an associated event may be generated at the wall switch (shown in
For example, an event may include a time component 351 of 17:12:02, an identification component 352 of “WALLSWITCH102,” and a state component of “ON.” This may represent an event created at 17:12:02 on the smart box connected to the wall switch and may represent that the wall switch was toggled from ‘off’ to ‘on.’ Continuing this illustration, occurrence data describing such an event may be broadcast in an event report message from the smart box associated with the wall switch to any smart box within its broadcast range. The smart box associated with the floor lamp may receive the broadcasted event report message and process the included occurrence data to generate an event for processing with a learning algorithm as described below.
An event pattern may include one or more events obtained, generated, or otherwise encountered in a time window or sequence. For example, a particular event pattern may include a first event generated internally by a learning device (e.g., a smart floor lamp, etc.) and a second event obtained by the learning device in response to receiving a signal received from another device (e.g., a smart wall switch, etc.). As later described, event patterns may be trigger patterns, action patterns, correction patterns, or reward patterns. Regardless of which type, event patterns may be order-dependent, such that the order in which particular events are received constitutes a pattern. Alternatively, event patterns may be order-independent where the pattern is independent of the processing order for the events. For example, a first event (referred to as event A) may be obtained (e.g., generated based on received occurrence data) at time 0 and second and third events (referred to as event B and event C respectively) may be obtained simultaneously at a later time 1 (denoted as A:0, B:1, C:1). In an order-dependent pattern, the learning device may only recognize the pattern if event A is obtained first and events B and C are simultaneously obtained after event A (denoted as A:0, B:1, C:1). However, if event C is obtained at time 2 instead of time 1, then the pattern (A:0, B:1, C:2) may not equal the pattern A:0, B:1, C:1 because the event C was obtained at time 2 instead of time 1. Thus, the first pattern created by obtaining event C at time 1 (A:0, B:1, C:1) and the second pattern created by obtaining event C at time 2 (A:0, B:1, C:2) are different because the times for obtaining event C are different. In an order-independent pattern, the learning device may treat obtained events A:0, B:1, C:1 the same as obtained events A:0, B:1, C:2 because the time of C is not important so long as event C is obtained within the same predetermined time window as event A and event B. In other words, for order-independence, the same events merely need to be obtained within a particular time window. Time windows observed by smart boxes or learning devices are further described below with reference to
In some embodiments, multiple smart boxes or learning devices may generate patterns (e.g., trigger patterns and action patterns) and conduct actions based on a single event. For example a user may toggle the wall switch from ‘off’ to ‘on’ causing the wall switch to generate a single first event. Upon generating the first event, the wall switch may broadcast a related event report message wirelessly to all nearby learning devices. A first nearby learning device may be the floor lamp, for example, which may generate the first event based on the received event report message and convert it to a trigger pattern. In response to the trigger pattern, the floor lamp may generate an action pattern and activate the light based on the action pattern. Simultaneously, a nearby stereo may receive the same event report message and similarly generate the first event based on the received event report message, convert it to a trigger pattern, generate a different associated action pattern than the floor lamp, and play music based the different action pattern. Thus, a single broadcasted event report message related to the first event in this example caused the floor lamp to activate its light and the stereo to play music.
In some embodiments, multiple smart boxes may generate action patterns and conduct corresponding actions based on receiving multiple event report messages related to multiple individual events. For example, a user may toggle the wall switch from
Similarly,
As described above, a reflex may be stored information that indicates a predefined action that a smart box may take or initiate in response to detecting an associated trigger. As illustrated in
When a smart box obtains an event (or multiple events) matching a known trigger pattern of a known reflex, the smart box may generate the corresponding action pattern 404. A reflex may have a predetermined reward pattern and a predetermined correction pattern. If a smart box receives a reward pattern when it is allowed to learn, the smart box may increase a weighting (i.e., the trigger weight) on the association between the trigger pattern 402 and the action pattern 404. Once the association weighting exceeds a threshold amount, the smart box will may execute the action pattern in response to the trigger pattern. Similarly, a reflex 400 may have a predetermined correction pattern 408, and if a smart box receives a correction pattern when it is allowed to learn, the smart box may decrease the association weighting between the trigger pattern 402 and the action pattern 404. Processing of the correction pattern 408 may modify the association weighting enough times that the association weighting may drop below the threshold amount and the smart box will effectively learn not to perform the action pattern 404 in response to the trigger pattern 402. In this manner, the smart box may learn the association between a trigger pattern 402 and a corresponding action pattern 404, and unlearn undesired trigger/action associations. In various embodiments, the correction pattern 408 and/or the reward pattern 406 may be obtained based on data received by the smart box from another smart box device, such as a nearby device emitting event report messages in response to performing an action, receiving an input, etc.
In some embodiments, a method of enabling an “allowed to learn” state (or a learning mode) for the smart box may be used to associate a predefined action pattern 404 of a reflex 400 of the smart box with a trigger pattern. Such a learning mode may be an operational state of the smart box during which the smart box may be enabled to change trigger weights of the reflex 400. Once an obtained pattern is matched to a trigger pattern 402 of a known reflex 400, the reflex may enter the learning mode. In other embodiments, the smart box may enter the learning mode when the action pattern 404 is generated. In other embodiments, the smart box may enter a global learning mode or state, which may be independent of triggers (e.g., turning on a learning switch) and during which the smart box may change trigger weights for various reflexes or otherwise generate new reflexes based on obtained events. In various embodiments, a reflex 400 may include data indicating the status of its various modes, such as bits, flags, or other indicators indicating whether the reflex 400 is in an active monitoring mode, triggered mode, learning mode, etc.
A smart box may be configured with one or more reflexes with action patterns for predetermined, known capabilities of the smart box. Although the smart box may utilize multiple reflexes with different corresponding actions, in some embodiments, the smart box may not be configured to perform actions outside of the static set of known capabilities or actions of the smart box, such as action patterns indicated in data provided by a manufacturer. Thus, the smart box may be configured to generate new reflexes with unknown triggers correlated to known actions, but may not be configured to generate new reflexes with actions that are not predefined.
As an illustration, a stereo learning device (or a stereo coupled to a learning device or smart box) may be configured with predetermined actions for setting a volume level to any value in a finite range of volume level values (e.g., 0-10, etc.), activating a radio (or radio tuner) ‘on’, deactivating the radio (or radio tuner), setting a radio station to any value in a finite range of radio station values (e.g., 88.1-121.9, etc.), setting a frequency modulation (FM) configuration or an amplitude modulation (AM) configuration, etc. The stereo learning device may store reflexes for each of these predetermined actions with various trigger patterns. For example, the stereo learning device may store a first reflex with an action pattern that sets the radio station to a first value (e.g., 92.3 FM) and a trigger pattern of a lamp ‘on’ event, a second reflex with an action pattern that sets the radio station to a second value (e.g., 101.5 FM) and a trigger pattern of a wall switch ‘on’ event, a third reflex with an action pattern that sets the volume level to 8 and a trigger pattern of the lamp ‘on’ event, etc.
Patterns may be created from one or more events (e.g., time component, device component, etc.) obtained at a smart box, such as events generated based on occurrence data obtained by a sensor (e.g. a light sensor, a switch vision sensor, etc.) and/or one or more events generated based on occurrence data received by the signal receiver 142. Events may be stored in memory 138 and used by the event recorder 206 to create or recognize patterns. Prior to evaluating events to create or recognize patterns, a filter may be applied to events to reduce the set of events that may be considered. For example, a floor lamp smart box may ignore events related to event report messages from a stereo. As an alternate example, the stereo may ignore events obtained or generated after sometime of day, such as 11:00 PM. Once a smart box generates a pattern of events, it may determine whether the pattern matches any known trigger patterns that correspond to a stored reflex.
If an identified pattern matches a stored trigger pattern in a reflex and the related trigger weight is equal to or above a particular threshold, its paired action pattern may be generated. A current trigger weight (Wi) for a certain reflex (Reflexi) may be calculated based on the following equation:
W
i=(Σk=0nmk,ixk,isk,i)+bi;
where i is a reflex counter or identifier, n is the number of events associated with a trigger pattern of the reflex, k identifies a counter for individual events in the trigger pattern of the reflex, m is an event match indicator for an individual event in the trigger pattern of the reflex, x is a match weight associated with the individual event in the trigger pattern of the reflex, s is a scale factor applied to the individual event in the trigger pattern of the reflex, and b is a bias for an entire weight match applied to the individual event in the trigger pattern of the reflex. Thus, the current trigger weight, Wi, of a Reflexi equals the sum of the event match (m) multiplied by the match weight (x) and the scale factor (s) plus the bias b in the trigger pattern associated with Reflexi. In some embodiments, match weights (x) may be adjusted by gains associated with their respective events, and as described in this disclosure, gains may be set based on whether a learning device is within a critical period or steady state period. In some embodiments, smart boxes may normalize values from 0.0 to 1.0. Further, in some embodiments, the event match indicator for an event (m) may be a floating value between 0.0 and 1.0 that may indicate whether the event was matched perfectly or not. (i.e., an event match value of 1.0 may represent a perfect match and an event match of 0.0 may indicate a complete mismatch).
As an illustration, if an identified pattern of a single event matches a known trigger pattern for a certain reflex (Reflexi), then the event match indicator (m) for the single event may be set to 1. Assuming the match weight (x) for the single event is set to 1 based on an associated gain value, the scale factor (s) is also set to 1, and the bias (b) for Reflexi is set to 0, then the new or current trigger weight Wi for the Reflexi may be equal to 1. If the same pattern is received again, then the match weight (x) may be adjusted by the current gain associated with the reflex, resulting in an increase in a subsequent, new trigger weight (Wi) that may be greater than the trigger weight threshold. Thus, the new trigger weight (Wi) may increase or decrease. For example, receiving the same trigger pattern a second time may increase the trigger weight (Wi) to 1.5 assuming that mk,i is set to 1, xk,i is adjusted to 1.5, sk,i is set to 1, and bi is set to 0. Under the same conditions, if the identified pattern does not match a known trigger pattern, then m may be equal to 0 resulting in a new trigger weight Wi also equal to 0.
As an additional illustration, a stereo (e.g., stereo 106 as described above in
The following tables illustrate exemplary properties of the equation with respect to the first reflex of the stereo (i.e., Ri). For the purpose of the following examples and tables, the action pattern (i.e., turning the stereo on and playing music) of the first reflex may be triggered when the trigger weight of the first reflex (i.e., Wi) is greater than or equal to a trigger threshold value of 1.5, a condition that may occur in response to the stereo receiving at least one of the first event and the second event. The first event may be event k=0 and the second event may be k=1. Further, except for the match indicator for various events (mn,i), it should be appreciated that the various values in the following properties may be predefined, such as set by a manufacturer, developer, or user. For example, the match weight for an event may be set by a manufacturer or may be based on previous events encountered at a smart box.
As shown in the exemplary properties of Table A above, in one scenario, only the first event (i.e., k=0) may be received by the stereo. Thus, the smart box of the stereo may set the event match indicator for the first event (m0,i) to 1.0 (i.e., there is a match for the first event) and the event match indicator for the second event (m1,i) to 0.0 (i.e., there is no match for the second event). The trigger weight of the first reflex may be computed by summing the sub-weight calculation for each event, such that the sub-weight of the first event computes to 1.0. In other words, (m0,i*x0,i*s0,i)+bi=(1.0*1.0*1.0)+0.0=1.0. As there is no second event, the event match indicator for the second event (m1,i) may be 0.0, and thus the sub-weight calculation for the second event may be 0.0. In other words, (m1,i*x1,i*s1,i)+bi=(0.0*1.0*1.0)+0.0=0.0. Accordingly, the total trigger weight of the first reflex (Wi) is 1.0 (i.e., 1.0+0.0), which is less than the trigger threshold value of 1.5. Thus, with only the first event received, the action pattern of the first reflex may not be triggered (e.g., the stereo may not activate its radio).
As shown in the exemplary properties of Table B above, in another scenario, both the first event (i.e., k=0) and the second event (i.e., k=1) may be received by the stereo. Thus, the smart box may set the event match indicator for the first event (m0,i) to 1.0 (i.e., there is a match for the first event), and the event match indicator for the second event (m1,i) to a non-zero value. However, in some cases, the second event may not be matched exactly, and thus the match indicator for the second event (m1,i) may be set to 0.8 (i.e., there is at least a partial match for the second event). The value of 0.8 for the event match indicator for the second event (m1,i) may indicate that the second event match was an imperfect match for a system that normalizes values from 0.0 to 1.0; where 1.0 represents a perfect match for the event match value.
As described above, the trigger weight (Wi) may be computed by summing the sub-weight calculation for each event, such that the sub-weight of the first event computes to 1.0. In other words, (m0,i*x0,i*s0,i)+bi=(1.0*1.0*1.0)+0.0=1.0. Further, the sub-weight of the second event computes to 0.8. In other words, (m1,i*x1,i*s1,i)+bi=(0.8*1.0*1.0)+0.0=0.8. Accordingly, the total trigger weight of the first reflex (Wi) may be 1.8 (i.e., 1.0+0.8), which is greater than the trigger threshold value of 1.5. Thus, with both the first event and the second event obtained at the smart box, the action pattern of the first reflex may be generated, causing an action to be performed (e.g., the stereo may activate its radio and play music, etc.). In some embodiments, the action pattern of the first reflex may be generated and cause an action to be performed in response to the calculation of any total trigger weight of the first reflex (Wi) that is greater than or equal to the trigger threshold value (e.g., 1.5).
In some embodiments, based on the match weights for various events, the smart box may be configured to perform actions in response to obtaining a single event. For example, the stereo smart box may be configured to activate its radio functionality in response to only receiving a signal indicating someone has sat in the recliner (i.e., the action pattern may be triggered by a presence sensor event associated with the recliner). As shown in the exemplary properties in Table C above, the first event may not be obtained (i.e., m0,i=0.0), the second event may be obtained (i.e., m1,i=0.8), and the match weight for the second event (x1,i) may be set to a value of 2.0. Due to the higher match weight for the second event, the radio of the stereo may be activated when only the second event is obtained at the stereo. In other words, the trigger weight for the first reflex may be greater than 1.5 based only on obtaining the second event (i.e., ((m0,i*x0,i*s0,i)+(m1,i*x1,i*s1,i)+bi=((0.0*1.0*1.0)+(0.8*2.0*1.0))+0.0=1.6.
In some embodiments, when imperfect event matching is likely, such as in a noisy RF environment, scale factors may be adjusted such that reflexes may be triggered even when matching may be low. For example, as shown in Table D above, the scale factor for the first event (s0,i) and the scale factor for the second event (s1,i) may be increased to a value of 2.0 in order to enable trigger weights above the 1.5 threshold value, even when matching indicators are less than ideal (e.g., less than 1.0, less than 0.8, etc.). In other words, the stereo may activate its radio to play music in response to receiving both the first event and the second event with less than ideal matching indicators (e.g., 0.7 and 0.6, respectively) and calculating a trigger weight of 2.6 for the first reflex (i.e., ((m0,i*x0,i*s0,i)+(m1,i*x1,i*s1,i)+bi=((0.7*1.0*2.0)+(0.6*1.0*2.0))+0.0=2.6.
In some embodiments, bias values for trigger weight calculations may be adjusted in order to cause action patterns to be triggered in response to a smart box obtaining a single event. For example, as shown in Table E above, the bias (bi) may be set to 1.0, which allows either the first event or the second event to individually cause the stereo to activate its radio via the first reflex. In other words, the action pattern may be triggered when only the second event is obtained (i.e., ((m0,i*x0,i*s0,i)+(m1,i*x1,i*s1,i))+bi=((0.0*1.0*1.0)+(0.8*1.0*1.0))+1.0=1.8) or when only the first event is obtained (i.e., ((m0,i*x0,i*s1,i)+(m1,i*x1,i*s1,i))+bi=((0.9*1.0*1.0)+(0.0*1.0*1.0))+1.0=1.9).
For the purposes of illustration, at time t=t0, the floor lamp may be considered to be in the monitor mode 506 with respect to all reflexes. The floor lamp may receive an event report message 502, such as via its signal receiver 142. For example, a user may toggle the wall switch from ‘off’ to ‘on’. In response, the wall switch may record the toggle as an event with a sensor encoder 134 (shown in
At t=tTrigger, the event report message 502 may be received by the floor lamp. The floor lamp may determine that an event generated based on the event report message 502 matches a trigger pattern of a reflex, and may enter the triggered mode 508 with respect to the matched reflex. During the triggered mode 508, the floor lamp may continue to search for other events to determine whether a reward and/or correction pattern is present to enable learning or unlearning, respectively.
At t=tResponse, the floor lamp may generate the event 514 associated with an action pattern of the matching reflex, which may activate a motor driver 140 to cause an action, such as turning on the light 124 of the floor lamp (shown in
At t=tResume Monitor, the floor lamp may leave the triggered mode 508 and re-enter the monitor mode 506 in which the floor lamp may search for and receive new event report messages.
As
The floor lamp may generate events of trigger patterns for different reflexes at different times, which may cause the floor lamp to enter triggered mode with respect to one reflex at a different time than the other triggered mode for the other reflex. Assuming the triggered modes of each reflex overlap the same time period (e.g., 5 seconds), the floor lamp may exit the triggered mode with respect to the first reflex but remain in the triggered mode with respect to the second reflex. Eventually, the floor lamp may exit the triggered mode with respect to each reflex and return to the monitor mode with respect to each reflex.
At t=t0, the floor lamp may start in the monitor mode 606 with respect to ReflexF1. ReflexF1 may include a trigger pattern (referred to as MD2), an action pattern (referred to as MD3), a reward pattern (referred to as MD4), and a correction pattern (referred to as MD5). The floor lamp may monitor generated events for patterns that match the trigger pattern of ReflexF1 (MD2).
At t=tMd1-on, the new wall switch may be switched from ‘off’ to ‘on’ generating an event and causing related occurrence data (referred to as “occurrence data 1”) to be broadcast by the wall switch in an event report message received by the floor lamp. The occurrence data from the event report message from the new wall switch may be used by the floor lamp to generate an event that may be combined with one or more events or may individually be used to create a pattern (“MD1”).
At t=tMd1-done, the floor lamp may receive the event report message with “occurrence data 1,” generate a related event, and convert it (and possible other events stored in a buffer) into a pattern known as pattern “MD1”. At this time, the floor lamp may place pattern MD1 on the event bus for further processing or temporary storage in memory. The floor lamp may determine that pattern MD1 does not match any known trigger patterns of known reflexes of the floor lamp and thus may continue to operate in the monitor mode 606.
At t=tMd2-on, the lamp switch may be turned from ‘off’ to ‘on’ and, in response, the floor lamp may generate an event based on occurrence data related to the state change (referred to as “occurrence data 2”). Simultaneously, the floor lamp may combine the event generated from the “occurrence data 2” with other events collectively processed as a pattern MD2 and place the pattern MD2 on the event bus for temporary storage in memory.
At t=tTrigger, the floor lamp may match the pattern MD2 to the trigger pattern of ReflexF1. The floor lamp may then enter the triggered mode 608 for ReflexF1 because pattern MD2 matches the trigger pattern of ReflexF1. In some embodiments, the floor lamp may complete an internal transmission and convert the event generated from the “occurrence data 2” into the pattern MD2 at t=tTrigger.
At t=tAction, the floor lamp may generate the action pattern for ReflexF1 (MD3) associated with the known trigger pattern for ReflexF1 (MD2) that is located on event bus or stored in the floor lamp's memory. The generation of the pattern MD3 may cause a motor driver connected to the floor lamp to turn on a light.
At t=tNewReflex, a new reflex (referred to as “ReflexF2” or “F2”) is created because there is no existing reflex with a trigger pattern matching the pattern MD1. The only known trigger pattern is MD2 associated with ReflexF1. In creating ReflexF2, the floor lamp may copy the action pattern, the reward pattern, and the correction pattern associated with the ReflexF1 into the new reflex, and may assign the pattern (MD1) received on the timeline 601 to the new reflex as its trigger pattern. The weights associated with the copied patterns may be adjusted when copied to the new reflex. Thus, the new reflex (ReflexF2) may have a trigger pattern equal to pattern MD1 and related to the occurrence data received from the new wall switch (“occurrence data 1”), an action pattern equal to pattern MD3 associated with turning the floor lamp on, a reward pattern equal to pattern MD4, and a correction pattern equal to pattern MD5. In some embodiments, when the floor lamp may be configured to perform more than one action (e.g., turn on, turn off, etc.) and thus utilize at least two reflexes (i.e., at least one reflex per action), then new reflexes created in response to detecting unknown patterns may be copied from an existing reflex in its triggered mode. In other words, in order to determine which existing reflex to copy from when creating a new reflex, the floor lamp may perform operations to correlate events (or patterns of events) with known actions of reflexes in their triggered mode (i.e., patterns for a new reflex may be copied from a pre-existing reflex whose action pattern is encountered within a time window of the unknown pattern/event).
At t=tReward another component may generate events that match a reward pattern, such as pattern MD4 known as the reward pattern for ReflexF1. For example, a motor driver may generate an event equal to pattern MD4 when the light of the floor lamp turns on (shown in
While in a learning-enabled mode, if a reward pattern (MD4) is matched, then reward gains may be applied (e.g., increasing the trigger weight, etc.). In some embodiments, although the match weight (x as described above) is typically modified while in a learning-enabled mode, any parameter or value in the equation may be adjusted while in a learning-enabled mode. In other words, increasing or decreasing the trigger weight of a reflex may include adjusting any parameter in the trigger weight equation.
However, if the correction pattern (MD5) is matched, then the correction gains may be applied (e.g., decreasing the trigger weight). In some embodiments, the reward pattern or the correction pattern may be generated by an additional occurrence, such as an input or a button that the user may activate in order to provide feedback that the response was as desired (or not desired). For example, after the floor lamp turns its light on, a user may press a button on the floor lamp, which generates a reward pattern. Based on the reward pattern, the floor lamp may increase the trigger weight of the related reflex.
At t=tResume Monitor, the floor lamp ends its triggered mode 608 for ReflexF1 and returns to the monitor mode 606. In some embodiments, the floor lamp may subsequently receive pattern MD1, which may cause the floor lamp to activate its light based on a triggered action of ReflexF2.
In some embodiments, a new reflex may be generated regardless of the order in which various occurrence data is received or obtained by the floor lamp. In other words, an unknown trigger pattern (e.g., MD1) could be received and used before, during, and after a trigger window and thus cause the creation of a reflex independent of the order of receiving occurrence data. For example, if the “occurrence data 1” is received and used to generate the pattern MD1 after the floor lamp has entered its triggered mode 608 for ReflexF1 (i.e., after “occurrence data 2” is received and MD2 has been obtained), the floor lamp may still create ReflexF2, as the MD1 may still have occurred within a time window relative to the triggered mode 608.
At t=t0, the wall switch may generate an event and broadcast an event report message with occurrence data related to the event. The floor lamp, in the monitor mode 706, may receive the event report message by t=tMD1-Rx.
At t=tMD1-Rx, the floor lamp receives the entire event report message with the occurrence data, generate an event in response, and transfers it to the event recorder, which may convert the event into pattern MD1 and place it on a event bus (as shown in
At t=tTrigger, the floor lamp may process pattern MD1 and determine that it matches a known trigger pattern associated with ReflexF2. Thus, the floor lamp may enter the triggered mode 708 with respect to ReflexF2 where floor lamp may learn or unlearn with respect to ReflexF2.
At t=tAction, the floor lamp may generate the action pattern (MD3) associated with ReflexF2, which is put on the event bus. The motor driver may retrieve the action pattern (MD3) from the event bus and conduct an action associated with the generated action pattern (e.g., turn on the light of floor lamp).
At t=tReward, a reward pattern (MD4) associated with ReflexF2 may be generated from another component. For example, the generated action pattern (MD3) may cause a motor driver to turn on the floor lamp. When the floor lamp turns on, the motor driver may receive feedback or a sensor encoder may sense a change in state on the lamp, thus generating pattern MD4. Pattern MD4 may be subsequently stored in event pattern storage. Pattern MD4 may match the reward pattern of ReflexF2, and as a result the weights associated with the ReflexF2 trigger pattern (MD1) may be increased.
In some embodiments, once the trigger weight of a reflex reaches a maximum level, the trigger weight may not be further adjusted, allowing system resources to be used elsewhere. Such a maximum level may be utilized to limit the dynamic range of the weight calculations or to reduce the amount of RAM included within learning devices. For example, when a less dynamic range of trigger weights are used for a reflex (e.g., a smaller range in between a minimum and maximum trigger weight), less RAM may be used in learning devices (e.g., 8-bits instead of 16-bits).
In some embodiments, the memory of the floor lamp may be of a size such that it may only store a limited number of patterns and/or reflexes. In such a case, if a trigger weight of a stored reflex reaches a minimum weight value (e.g., a ‘discard threshold’), the trigger weight may be considered so low that it may likely never trigger a reflex. In such a case, the floor lamp may re-use (or reclaim) the memory allocated to that reflex for new reflexes. Thus, setting a lower limit for correcting a reflex with a low trigger weight may allow the memory to devote storage for other patterns and/or reflexes. In other embodiments, when there are limited resources for storing new reflexes, the floor lamp may reallocate memory from the most infrequently used or lowest likely to be used (via weight properties) to new reflexes without using a minimum or “discard” threshold (i.e., the floor lamp may simply replace the most useless reflexes).
At t=tCorrection, a different component may generate a correction pattern (MD5). For example, if the floor lamp is turned off within the triggered mode 708, a sensor encoder may convert this change in state to an event, which may be passed to the event recorder to create a correction pattern MD5. Pattern MD5 may be matched to the correction pattern of ReflexF2 (which is in the triggered mode 708), and as a result the trigger weights may be reduced to weaken the association between the trigger pattern (MD1) and the action pattern (MD3) of ReflexF2.
At t=tResume Monitor, the floor lamp may exit the triggered mode 708 associated with ReflexF2, and the floor lamp may return to monitor mode 706. The triggered mode 708 may end simply because it has been timed out. For example, a triggered mode 708 may only last for ten seconds, so after operating in the triggered mode 708 for ten seconds, the floor lamp may exit the triggered mode 708 with respect to ReflexF2 and may enter a corresponding monitor mode 706.
The critical learning period 801 may be typically associated with the initial state of the learning device. This may be a time in which training the initial behavior of the learning device would be more beneficial to the user. Initial dynamic reflexes are likely to be created in this state; meaning that various gain values associated with the critical learning period 801 (referred to as “Gain Set 1” in
The steady state learning period 802 may occur when a particular device has been initially trained, and additional training is allowed but is intended to be more difficult. Gains associated with the steady state learning period 802 (referred to as “Gain Set 2” in
The relation of the gains associated with the critical learning period 801 (“Gain Set 1”), and the gains associated with the steady state learning period 802 (“Gain Set 2”) may be illustrated with the following equation:
Gain Set 1>Gain Set 2
In other words, a learning device using the above equation may learn more quickly with Gain Set 1 than Gain Set 2.
In some embodiments, each gain set may have individual gains or weights associated with the trigger, reward, and correction pattern of a reflex at different stages of operation. Two or more gain levels may be used to adjust the gains closer to a critical period and a steady state period. For example, there may be a third gain set, which may be a hybrid between the critical period and the steady state period (e.g., less repetition is needed to learn). As the gains are adjusted, the weights associated with a particular pattern may be adjusted to determine matches within the system.
Whether a particular reflex is dynamic or static may affect the gains and learning associated with the learning device. A particular learning device may have a built-in static reflex, which may not be adjusted. For example, the floor lamp may have a built-in reflex incapable of being re-weighted regardless of encountering related reward patterns or correction patterns. In other words, learning devices may not nullify (or “forget”) static reflexes through the use of weight adjustments (e.g., correcting). However, in contrast, dynamic reflexes may be created spontaneously and may be adjusted over time. For example, the floor lamp may adjust the weights of a dynamic reflex (e.g., ReflexF2 as illustrated above) over time such that no action of the floor lamp may be performed corresponding to a trigger pattern associated with the wall switch. In other words, a learning device may lower the trigger weight of a reflex related to the association between a trigger pattern (e.g., occurrence at a wall switch) and an action pattern (e.g., turning on the floor lamp) such that the trigger weight is below a threshold and thus the action may not be performed. However, in some embodiments, dynamic reflexes may be converted to static reflexes such that the association may not be forgotten. In some embodiments, dynamic reflexes may be given a rigid state such that it is difficult to change the trigger weight of a reflex having an association between an action and a trigger, thus making such dynamic reflexes more persistent.
At time t=t0, the floor lamp may be in the monitor mode 906 with respect to ReflexF2. In the monitor mode 906, the floor lamp may monitor for incoming signals related to events matching the trigger pattern ReflexF2. During the monitor mode 906, the floor lamp may encounter or obtain an event corresponding to trigger pattern MD1. For example, a new wall switch, which may be identical to the first wall switch, may send an event report message with occurrence data to the floor lamp when the new wall switch toggles from ‘off’ to ‘on’, and the floor lamp may then generate a trigger pattern MD1 based on the received event report message and occurrence data.
At time t=tNoAction1, the floor lamp may process the trigger pattern MD1 for ReflexF1 and ReflexF2. As previously discussed, MD1 may only be associated with ReflexF2, thus floor lamp may enter the triggered mode 908 with respect to ReflexF2. Since ReflexF2 has a current trigger weight at a first trigger weight level 921 that is below the trigger weight threshold 925 at t=tNoAction1, the floor lamp may not generate the action pattern for ReflexF2 (e.g., MD3). However, shortly thereafter, the floor lamp may generate trigger pattern MD2 after receiving another event report message with occurrence data corresponding to the new wall switch. For example, the new wall switch may toggle from ‘off’ to ‘on’ and send a related event report message to the floor lamp, causing the floor lamp to generate the trigger pattern MD2 based on the event report message. As trigger pattern MD2 corresponds to ReflexF1 and the trigger weight is above its trigger threshold, the floor lamp may generate action pattern MD3. The floor lamp may subsequently generate a corresponding action event that results in the lamp turning on its light. Once the light turns on, the change in state may be recorded by a sensor encoder, which creates an associated event and generates the reward pattern MD4.
At time t=tWeightAdjust1, reward pattern MD4 may be processed to adjust the trigger weights for both ReflexF1 and ReflexF2. While in triggered mode 908 with respect to ReflexF2, the floor lamp may determine that pattern MD4 matches the reward pattern of ReflexF2, and may increase the trigger weight of MD1 and ReflexF2. The new trigger weight is at a second trigger weight level 922, which is still below the trigger weight threshold 925. After the triggered mode 908 times out, the floor lamp may enter the monitor mode 906 again.
The process of encountering events and generating their corresponding patterns MD1, MD2, MD3 (or MD3′), and MD4 may repeat resulting in adjusting the trigger weight of ReflexF2 to increase above the trigger weight threshold 925 to a third trigger weight level 923 at t=tWeightAdjust2.
At any time after adjusting the trigger weight of ReflexF2 above the trigger weight threshold 925, the floor lamp may encounter an event corresponding to the pattern MD1, which may result in the generation of action pattern MD3′ without the need of encountering pattern MD2 to trigger ReflexF1. For example, before the floor lamp may have only turned on when it generated pattern MD2 corresponding to an ‘on’ event of the new wall switch. Now the wall switch may send an event report message including occurrence data that may result in the generation of an event corresponding to pattern MD1 to the floor lamp and thus in the floor lamp being triggered to turn on its light via ReflexF2.
At time t=t0, the floor lamp may monitor for events in a monitor mode 1006. During the monitor mode 1006, the floor lamp may encounter a trigger event corresponding to trigger pattern MD1. For example, a new wall switch may broadcast an event report message with occurrence data related to an ‘on’ event and corresponding to pattern MD1 because the new wall switch was toggled from ‘off’ to ‘on’. As the event is received, the floor lamp may generate the corresponding trigger pattern.
At time t=tTriggered1, the floor lamp may receive the event report message related to the on-event and generate the pattern MD1. The floor lamp may determine that the pattern MD1 is a known trigger pattern corresponding to ReflexF2 and thus may enter the triggered mode 1008 with respect to ReflexF2. Shortly thereafter, the floor lamp may determine that the first trigger weight level 1023 for ReflexF2 is above trigger weight threshold 1025 and may generate an action pattern MD3, which results in an action event and a physical action of the floor lamp turning on its light. The floor lamp may also encounter an event corresponding to a correction pattern MD5 while in the triggered mode 1008. For example, the floor lamp may generate a correction pattern MD5 upon encountering an event when a user presses a separate correction button on the floor lamp (e.g., a button labeled “Correction”). A user may press this button to send a correction event to the floor lamp and in response the floor lamp may generate the correction pattern MD5. In an alternative example, the floor lamp may generate a correction pattern when a user manually turns off the floor lamp within a brief time window of a previous trigger pattern. The opposite input of a previous trigger pattern may correspond to a correction pattern and the floor lamp may learn to disassociate trigger patterns and action patterns.
At time t=tCorrection1, the floor lamp may determine that the correction pattern MD5 matches the correction pattern of ReflexF2. Thus, the floor lamp may reduce the trigger weight associated with ReflexF2 to a second trigger weight level 1022. The second trigger weight level 1022 is still above the trigger weight threshold 1025, thus the floor lamp may still activate its light. Eventually, the triggered mode 1008 ends due to time constraints and the floor lamp may enter the monitor mode 1006 again.
While in the monitor mode 1006, the floor lamp may encounter a second trigger event and generate a second trigger pattern MD1. For example, the new wall switch may again be toggled from ‘off’ to ‘on’. At time t=tTriggered2, the floor lamp may determine that the second pattern MD1 matches the known trigger pattern of ReflexF2 and may enter triggered mode 1008 with respect to ReflexF2. Since ReflexF2 currently has a second trigger weight level 1022 above trigger weight threshold 1025, the floor lamp may generate action pattern MD3 and the associated mechanical action (e.g., turn on the light). While the floor lamp is in the triggered mode 1008 with ReflexF2, the floor lamp may again encounter a correction event from the correction button and generate the correction pattern MD5. Since pattern MD5 corresponds to ReflexF2, at time t=tCorrection2, the trigger weight is reduced to a third trigger weight level 1021, which is below the trigger weight threshold 1025. Thus, if the floor lamp encounters another trigger event and generates another trigger pattern MD1 at time t=tTriggered3, the floor lamp may not generate a corresponding action pattern MD3 in a triggered mode 1008. In other words, the floor lamp may have effectively forgotten the trigger action association of ReflexF2 and may not activate its light upon generating trigger pattern MD1 in the future (or at least until retrained to respond to that manner to the trigger pattern).
In some embodiments, trigger weights below their association trigger weight threshold may continually lowered in response to the floor lamp entering its trigger mode without encountering a reward pattern. For example, in
In some embodiments, the floor lamp may remove ReflexF2 immediately or at sometime after its trigger weight is below the trigger weight threshold 1025 and there is a memory shortage. Thus, if the floor lamp detects the trigger pattern (MD1) of ReflexF2 after ReflexF2 has been deleted, the floor lamp may create a new reflex with pattern MD1 as its trigger pattern assuming the other conditions are met (e.g., having a reward present during the triggered mode). In some embodiments, the floor lamp may remove a reflex that has a trigger weight above its associated threshold due to memory shortages (e.g., reaching a memory limit for stored reflexes). For example, when the floor lamp encounters a new trigger pattern within a triggered mode but has no available storage in local memory, the floor lamp may remove a stored reflex that has a trigger weight above a trigger threshold but that is not often used, least likely to be used, and/or has the lowest trigger weight of all reflexes with trigger weights exceeding their respective trigger weight thresholds.
In block 1102 the floor lamp may obtain an event. For example, the floor lamp may receive an event report message including occurrence data over a RF transmission from the wall switch and, based on the data in the event report message, the floor lamp may generate the event as a data structure as described above with reference to
In determination block 1104, the floor lamp may determine whether an event filter applies. Event filters may include time filters, type filters, device event filters, etc. In response to determining that an event filter applies (i.e., determination block 1104=“Yes”), the floor lamp, may discard the event from further processing in block 1106, and continue to monitor for new incoming signals in block 1102. In some embodiments, if the event filter is a time-based filter, there may be a preset schedule to discard events during the day. For example, the stereo may have a time filter that it will ignore obtained events from the hours of midnight to 10 AM. In another example, an event filter at the floor lamp may simply ignore all obtained events from the stereo. In a further example, the stereo may ignore obtained events associated with a particular user. In some embodiments, a wall switch may receive a User ID input (e.g., fingerprint data, a pass code, nearby mobile device data from Bluetooth or Near Field Communication (NFC), etc.) and include that User ID in the occurrence data within an event report message. A father who owns a stereo may not want anyone other than him to turn on his stereo with wall switch. Thus, the stereo may discard all obtained events if they do not contain the father's user ID, thereby preventing others from turning on the stereo with the wall switch. However, if an event filter does not apply (i.e., determination block 1104=“No”), the floor lamp may store the event in a buffer located in memory 138 (shown in
Assuming an event filter does not apply, the floor lamp may store the event in the buffer located in memory 138 in block 1108. The event may be stored in a buffer to facilitate generating a pattern at event recorder 206 while the floor lamp is in monitor mode. In other words, the floor lamp may perform buffering of events while in monitor mode. Although not shown, the floor lamp may buffer events in memory for a particular period of time (e.g., 5-10 seconds) and then discard the events to make room for new events.
In block 1110, the floor lamp may generate a pattern based on the event residing in the buffer. In some embodiments, the floor lamp may generate a pattern based on multiple events residing in the buffer, such as by retrieving and combining various events buffered in memory. For example, the floor lamp may have generated a pattern based on two events generated based on event report messages received when two different wall switches are turned to the ‘on’ position. Patterns may be generated by one of four ways: (1) based on the time-ordered sequence of events; (2) reducing multiple events to a singlet; (3) heuristics; and (4) removing time from events in pattern generation.
When generating a pattern based on a time-ordered sequence of events, the time the event is generated or otherwise obtained may matter. Thus, if an event is not created within a certain time window, the floor lamp may not generate a pattern based on the event. For example, the floor lamp may have a trigger pattern equivalent to an ‘on’ event related to the wall switch and an ‘on’ event related to the stereo. If the floor lamp obtains the ‘on’ event related to the wall switch within the time window but the ‘on’ event related to the stereo is obtained outside of the time window, then the floor lamp may not recognize the trigger event. In some embodiments, a pattern may only be generated if an event A is obtained before event B. For example, if the floor lamp obtains the stereo ‘on’ event prior to the wall switch ‘on’ event, the floor lamp may not recognize these events as a trigger pattern because the floor lamp only accepts trigger patterns when the wall switch event is obtained first.
In some embodiments, multiple events may be reduced to a single event or a singlet. For example, the floor lamp may obtain two ‘A’ events at different times and then a ‘B’ event, which are stored in the lamp event buffer. The floor lamp may generate a pattern based on one ‘A’ event and one ‘B’ event, discarding the second ‘A’ event. Thus, a trigger pattern having two ‘A’ events and a ‘B’ event may be reduced to a trigger pattern having one ‘A’ event and one ‘B’ event. Since the ‘A’ event is repeated at a different time, the floor lamp may ignore the repeated event.
In some embodiments, the floor lamp may conduct a series of heuristic calculations to determine whether to disregard the event. Some of these heuristic calculations may simply include a counting mechanism. For example, the floor lamp may determine whether it has received the ‘A’ event three times (e.g., an ‘on’ event related to the wall switch), at which point the floor lamp may generate a corresponding pattern such as a trigger pattern based on a heuristic rule of receiving the three ‘A’ events equates to generating a trigger pattern.
In some embodiments, the floor lamp may disregard time when creating patterns from events. Disregarding time may coincide with the heuristic calculations. For example, if the floor lamp receives three ‘A’ events and one ‘B’ event in memory 138, the floor lamp may perform a series of heuristic calculations to determine whether to generate a pattern based on the events without a time window. Disregarding time may also include order-independence. For example, the floor lamp may create the same pattern regardless of whether it obtains an ‘A’ event followed by a ‘B’ event or a ‘B’ event followed by an ‘A’ event.
In determination block 1112, the floor lamp may determine whether to apply a pattern filter. This may be similar to the event filter described with reference to determination block 1104, which may include stored ignore patterns, time-based filters, device type filters, etc. The floor lamp may employ the pattern filter to remove a pattern from memory (e.g., a 32K memory, 64K memory, etc.) when the pattern falls below a threshold, such as a time threshold. In response to the floor lamp determining that a pattern filter applies (i.e., determination block 1112=“Yes”), the floor lamp may discard the pattern and refrain from further processing of that pattern in block 1113. In some embodiments, the floor lamp may filter patterns generated for recently conducted actions. For example, when the floor lamp turns on, the floor lamp may generate an action pattern from an event. If the action pattern was not ignored for a period of time, the floor lamp may try to process the action pattern as a trigger pattern to another action (e.g., to turn on the stereo). To avoid the creation of a new trigger-action association, the floor lamp may create a temporary ignore pattern filter in which the floor lamp ignores generated action patterns for a short period of time. After the floor lamp discards the pattern, the floor lamp returns to obtaining new events in block 1102. In some embodiments, the floor lamp may constantly obtain events in block 1102.
In some embodiments, the floor lamp may apply pattern filters if a trigger weight of the pattern or the corresponding reflex is below a low threshold value. By applying a pattern filter, the floor lamp may be able to remove patterns from its memory when the threshold value of a particular reflex is below a certain set value. The floor lamp may reduce the trigger weight of a reflex through the correcting process described throughout the application. Removing patterns may allow the floor lamp to conserve resources (e.g., memory) for the creation of new reflexes. In some embodiments, the floor lamp may be configured to utilize a predetermined, limited number of reflexes (e.g., 2 reflexes per lamp) so that users are less likely to get confused regarding the floor lamp's learned capabilities at any given time, regardless of the available local storage. Such limits to stored reflexes may also have the added benefit of improving performance, such as by improving pattern matching speeds by decreasing the number of patterns that may need to be compared due to fewer stored reflexes and patterns.
Referring back to determination block 1112, in response to determining that a pattern filter does not apply (i.e., determination block 1112=“No”), the floor lamp may determine whether the generated pattern matches a known pattern in determination block 1114. For example, the floor lamp may determine that the received event is within the time window of the time based filter. Thus, the floor lamp continues to process the event as a pattern. The floor lamp may determine whether the generated pattern is a known pattern of any type, such as a known trigger pattern, a known correction pattern, a known reward pattern, etc.
As an example, in determination block 1114, the floor lamp may determine whether a generated pattern corresponds to a known trigger pattern of a reflex, such trigger pattern ‘MD2’ for reflex ‘ReflexF1’ described above with reference to
However, in response to determining that the generated pattern does not match a known pattern (i.e., determination block 1114=No), the floor lamp may determine whether to create a new reflex in determination block 1116. For example, as described above in the scenario of
In response to the floor lamp deciding not to create a new reflex (i.e., determination block 1116=“No”), the floor lamp may discard the generated pattern in block 1113 and begin to monitor for new events in block 1102. In some embodiments, the floor lamp may be switched to a non-learning mode in which the floor lamp cannot learn new associations, thereby disabling its ability to create new reflexes. For example, the floor lamp may have previously learned to turn its light on/off when wall switch sends an event report message associated with an on/off event. A user may be satisfied with this simple on/off association and may disable any additional learning by the floor lamp. Thus, the floor lamp may not learn additional associations with occurrences (e.g., power on, etc.) at the stereo or any other learning device. In other embodiments, the floor lamp may have other considerations (e.g., not enough memory, triggered mode timed out, etc.) to keep it from learning a new reflex.
In response to the floor lamp deciding to create a new reflex (i.e., determination block 1116=“Yes”), the floor lamp may store the new pattern as a trigger pattern for a new reflex in block 1118. A new reflex may be created with a predetermined action pattern, reward pattern and a correction pattern. Thus, in block 1119, the floor lamp may copy to the new reflex the action pattern, reward pattern, and correction pattern from the reflex that is currently in its triggered mode. For example, as illustrated in
As previously noted, the floor lamp may obtain additional events while in the triggered mode, and such additional events may be associated with or correlated to a different trigger. The floor lamp may attempt to identify or match patterns based on these additional events to patterns of reflexes stored in memory. However, the patterns based on these additional events may not correspond to a known pattern of a stored reflex, and the floor lamp may decide to create a new reflex. In other words, the floor lamp may create a second reflex with a trigger pattern, action pattern, correction pattern, and reward pattern when patterns based on the additional events do not correspond to at least one of the trigger pattern, action pattern, correction pattern, and reward pattern associated with a known reflex.
The floor lamp may determine whether the trigger weight of the reflex of the matching pattern is equal to or above the trigger threshold in determination block 1204. Continuing with the example of
In some embodiments, the floor lamp may be configured to generate a limited number of actions when in the triggered. For example, the floor lamp may only generate one action during any one triggered mode, regardless of the number of trigger patterns received during that triggered mode.
In optional block 1217, the floor lamp may broadcast an event report message based on the generated action, such as a broadcast message including occurrence data indicating the generated action (or its resulting event). In response to the floor lamp determining that the matched trigger weight is not greater than or equal to the trigger threshold for the reflex (i.e., determination block 1204=“No”), or if the action is generated with the operations in block 1216 and a broadcast is made with the operations in optional block 1217, the floor lamp may perform the operations in determination block 1220 described below.
In response to the floor lamp determining that the generated pattern does not match a known trigger pattern (i.e., determination block 1202=“No”), the floor lamp may determine whether the floor lamp is allowed to learn in determination block 1206. For example, the floor lamp may have previously processed the trigger pattern (e.g., MD1) and is currently monitoring for generated reward patterns and correction patterns while in a triggered mode. Thus, the floor lamp may obtain a reward event and generate the corresponding reward pattern (e.g., MD4) shortly after receiving the trigger pattern and entering the activated trigger mode for the associated reflex.
In response to determining that the floor lamp is not allowed to learn (i.e., determination block 1206=“No”), the floor lamp may perform the operations in determination block 1220 described below. For example, the floor lamp may have a designated time window of five seconds after generating a trigger pattern to learn/unlearn a new action associated with the trigger pattern (e.g., MD1). As long as a reward pattern or correction pattern is generated within the five-second window, the floor lamp may learn/unlearn actions with the trigger pattern (e.g., MD1); however, the floor lamp may not learn new associations or unlearn old associations if the received reward/correction pattern is outside the five second time window. In another example, the floor lamp may not be able to learn simply because the associated reflex is in an unlearn state or the associated reflex is a static reflex which may not learn or unlearn.
However, if the floor lamp determines that it is allowed to learn regarding an action-trigger association of a reflex (i.e. determination block 1206=“Yes”), in determination block 1208, the floor lamp may determine whether the generated pattern matches a reward pattern. In some embodiments, the floor lamp may receive or generate a reward pattern within a learning time window. For example, a user may press a reward button on the floor lamp within five seconds of switching on the wall switch and turning on the floor lamp. By pressing a reward button on the floor lamp, it may generate a reward pattern (e.g., pattern MD4 as illustrated in
In some embodiments, the floor lamp may be allowed to learn based on whether it is in the monitor mode or the trigger mode. For example, when in monitor mode for a particular reflex, the floor lamp is not allowed to learn regarding that reflex; however, learning may be allowed when in the triggered mode of the reflex. In some embodiments, one or more reflexes may be allowed to learn due to other factors, such as the overall state or configuration of the floor lamp. For example, the floor lamp may be configured to disallow any learning due to a system setting, such as an active debug mode during which various reflexes may be tested.
If the floor lamp determines that the generated pattern matches a reward pattern (i.e., determination block 1208=“Yes”), in block 1212a the floor lamp may adjust the trigger weight of the associated reflex. In some embodiments, the floor lamp may adjust the trigger weight associated with the appropriate reflex by increasing the trigger weight. For example, if the floor lamp receives or generates a pattern (e.g., MD4) within a five-second learning time window of encountering a trigger pattern (e.g., MD1), the floor lamp may increase the trigger weight of the reflex of the trigger pattern. After the trigger weights are adjusted, in block 1214, the floor lamp may store the adjusted trigger weights in memory 138 and the floor lamp may perform the operations in determination block 1220 as described below.
In some embodiments, the floor lamp may optionally perform the operations in determination block 1210 after performing the operations in block 1212a. In other words, the floor lamp may be configured to evaluate both whether a reward pattern has been matched in determination block 1208 and whether a correction pattern has been matched in determination block 1210 in response to determining it is allowed to learn (i.e., determination block 1206=“Yes”), regardless of the determinations of determination block 1208. In other words, reward and correction matches may be checked in parallel by the floor lamp.
If the floor lamp determines that the generated pattern does not match a known reward pattern (i.e., determination block 1208=“No”), the floor lamp may check for a correction pattern match in determination block 1210. In some embodiments, the floor lamp may receive or generate a correction pattern within a learning time window. For example, a user may press a correction button on the floor lamp within five seconds of switching on the wall switch and turning on the floor lamp. By pressing the correction button, the floor lamp may generate a correction pattern (e.g., pattern MD5 as illustrated in
If the floor lamp determines that the generated pattern matches a known correction pattern (i.e., determination block 1210=“Yes”), the floor lamp may adjust the trigger weight in block 1212b. In some embodiments, the floor lamp may decrease the trigger weights after receiving a correction pattern within the learning time window. For example, the floor lamp may generate a correction pattern (e.g., pattern MD5) when a user turns the lamp switch 126 of the floor lamp to ‘off’ within five seconds of generating a trigger pattern (e.g., MD1) associated with an ‘on’ event of the wall switch. The floor lamp may match the generated pattern (MD5) as the correction pattern of ReflexF2 and reduce the trigger weight associated with ReflexF2. In block 1214, the floor lamp may store the adjusted weights in memory 138 and the floor lamp may perform the operations in determination block 1220 as described below. In other words, the floor lamp may adjust one or more trigger weights of the reflex when the at least one additional event corresponds to at least one of a correction pattern and a reward pattern associated with the reflex.
In response to the floor lamp determining that the generated pattern does not match a correction pattern (i.e., determination block 1210=“No”), or in response to the floor lamp determining that the matched trigger weight is not greater than or equal to the trigger threshold (i.e., determination block 1204=“No”), or in response to the floor lamp determining that is it not allowed to learn (i.e., determination block 1206=“No”), or in response to the floor lamp performing the operations of blocks 1217 or 1214, the floor lamp may determine whether to return to a monitor mode in determination block 1220, such as based on an expired duration since entering the activated triggered mode with the operations in block 1203. De-activating the triggered mode may activate the monitor mode associated with the reflex. In response to the floor lamp determining that it should return to the monitor mode (i.e., determination block 1220=“Yes”), the floor lamp may deactivate the triggered mode for the reflex in block 1222. In response to the floor lamp determining that it should not return to the monitor mode (i.e., determination block 1220=“No”) or when the operations of block 1222 have been performed, the floor lamp may continue obtaining events in block 1102 of method 1100 as described above with reference to
As an illustration based on the scenario shown in
The floor lamp may then perform the operations in blocks 1102, 1104, 1108, and 1110 as described above with reference to
If the floor lamp subsequently obtains the same event and generates the first pattern (e.g., pattern MD1) based on other data received from the wall switch, the floor lamp may process the first pattern with reference to the first reflex in the operations of blocks 1102, 1104, 1108, 1110, 1112, and 1114 as described above with reference to
Continuing with the illustration, the floor lamp may process the matched first pattern (MD1) from the new wall switch event and determine that the matched pattern is a trigger pattern match (i.e., determination block 1202=“Yes”) and may activate the triggered mode for the first reflex (e.g., ReflexF2). However, the trigger weight for the first reflex (e.g., ReflexF2) may be below its trigger threshold, in which case the floor lamp will not generate an action in block 1216 but instead may continue to monitor for other events/patterns. On the other hand, the floor lamp may encounter a different trigger event, such as an ‘on’ event from the lamp switch 126. The floor lamp may process the on-event from the lamp switch 126 through blocks 1102, 1104, 1108, and 1110 of method 1100 as described above with reference to
The embodiment methods described above with reference to
In this example, each of the floor lamp and wall switch may have a predefined reflex stored in the memory of their associated smart box. For example, the wall switch may have a predefined reflex, ReflexW, stored in memory that may include a trigger pattern, ‘WT’, an action pattern, ‘WA’, a correction pattern, ‘WC’, and a reward pattern, ‘WR’. The trigger pattern, WT, may correspond to a trigger event where a user toggles the wall switch from ‘off’ to ‘on’. When the user toggles the wall switch from ‘off’ to ‘on’, the wall switch may generate an event as well as broadcast an event report message including occurrence data related to the ‘on’ event. From the generated event related to the wall switch's ‘on’ event, a smart box included within or coupled to the wall switch may generate the trigger pattern WT. Initially, the action pattern, WA, may not correspond to a real life action such as toggling a switch. Instead, WA may simply be computer code ready to be assigned to future reflexes.
The correction pattern WC may correspond to a button on the wall switch labeled “Correction.” When a user presses the correction button, the wall switch may generate a correction event as well as broadcast another event report message with occurrence data indicating the correction event. From the generated correction event, the smart box associated with the wall switch may generate the correction pattern WC. The reward pattern, WR, may correspond to an event in which the user presses a reward button on the wall switch labeled “Reward.” When the user presses the reward button, the wall switch may generate a reward event as well as broadcast another event report message with occurrence data indicating the reward event. From the generated reward event, the smart box associated with the wall switch may generate the reward pattern, WR.
Similarly, the floor lamp may have a predefined reflex, ReflexF2, stored in memory that may include a trigger pattern, MD1, an action pattern, MD3, a correction pattern, MD5, and a reward pattern, MD4. The trigger pattern, MD1, may correspond to a trigger event in which the user toggles a lamp switch of the floor lamp from off to on. When the user toggles the lamp switch from ‘off’ to ‘on’, the wall switch may generate a trigger event as well as broadcast an event report message with occurrence data indicating the lamp's ‘on’ event. From the generated trigger event, a smart box included within or coupled to the floor lamp may generate the trigger pattern MD1. The action pattern, MD3, may correspond to an event in which the floor lamp turns its light from ‘off’ to ‘on’. The correction pattern, MD5, may correspond to an additional button on the floor lamp labeled “Correction” when the floor lamp is in a triggered mode. When a user presses the correction button, the lamp may generate a correction event as well as broadcast an event report message with occurrence data indicating the lamp's correction event. From the generated correction event, the smart box associated with the floor lamp may generate the correction pattern MD5. The reward pattern, MD4, corresponds to when the user turns on the floor lamp within the triggered mode, generating a reward event as well as broadcasting an event report message with occurrence data indicating the lamp's reward event. From the reward event, the smart box associated with the floor lamp may generate the reward pattern MD4.
With the wall switch and floor lamp initially configured in this manner, a user may train the floor lamp to turn on in response to the wall switch as follows. With the wall switch in the ‘off’ position and the floor lamp turned off, the user may turn on the wall switch and promptly turn on the floor lamp via manual operations (e.g., flipping switches on the devices). If the two actions are accomplished within a short time period (e.g., 5 to 10 seconds), the smart box associated with the floor lamp may begin to learn the turn-on action correlation by increasing a weight associated with the lamp-on reflex. Similarly, the user may teach the floor lamp to respond to the wall switch being turned off by turning off the wall switch and promptly turning off the floor lamp. Again, if the two actions are accomplished within a short time period (e.g., 5 to 10 seconds), the smart box associated with the floor lamp may begin to learn the turn-off action correlation by increasing a weight associated with the lamp-off reflex.
One such training cycle may not be enough (except in some embodiments in which a previously untrained smart box will learn a first reflex-event correlation in a single step), so the user may repeat the process of turning on the wall switch and promptly turning on the floor lamp, followed a while later by turning off the wall switch and promptly turning off the floor lamp. This series of steps may need to be repeated three or more times, depending upon the learning hysteresis configuration of the smart box associated with the floor lamp.
After two, three or more repetitions, the smart box associated with the floor lamp may have increased the weight associated with the lamp-on and lamp-off reflexes such that a subsequent toggle of the wall switch will cause the floor lamp to turn on or off accordingly. Thus, to train this desired correlation of the wall switch on/off events to the floor lamp on/off actions, the user may simply repeat the process until the floor lamp begins turning on in response to the user toggling the wall switch.
This series of actions by the user causes the following actions to occur within the smart boxes associated with the wall switch and the floor lamp. When the smart box associated with the wall switch senses the ‘off’ to ‘on’ toggle, the wall switch may generate a trigger event and associated occurrence data that may be broadcast in an event report message for all learning devices within its broadcast range (e.g., 100 feet) to receive. The smart box associated with the floor lamp, being within the broadcast range of the wall switch, may receive the event report message. Upon receipt, the floor lamp may generate a related event and determine whether an event filter stored in memory prevents further processing of the generated event. In a default state, the floor lamp may not apply a filter to the generated event, thus the floor lamp may store the generated event in a buffer. Based on the generated event, the floor lamp may generate the pattern MD2.
Initially, the floor lamp only has stored patterns associated with ReflexF2 in memory (e.g., MD1, MD3, MD5, and MD4). It is assumed for the purposes of this example that the floor lamp is already within a triggered mode for ReflexF2, such as in response to generating the trigger pattern for ReflexF2, pattern MD1, within the same time window as generating the pattern MD2. As the generated pattern MD2 does not match any pattern of the ReflexF2, the generated pattern MD2 may be considered an unknown pattern that may be used as a trigger pattern for a new reflex. The floor lamp may determine whether to create a new reflex with the unknown or unmatched pattern MD2. The floor lamp may have many different reasons to not create a new reflex. For example, the floor lamp may be in a learning prevention mode (e.g., a hold mode) or the floor lamp may be prohibited from creating reflexes from certain patterns associated with a particular device (e.g., the floor lamp will not create reflexes from patterns associated with the wall switch).
In this example the floor lamp is not prevented from creating a new reflex, so the floor lamp may create a new reflex, ReflexF1, with the unknown pattern, MD2, as its trigger pattern. The new reflex, ReflexF1, will include an action pattern, a correction pattern, and a reward pattern to be a complete reflex. Thus, the floor lamp may use patterns from the only known reflex in a triggered mode, ReflexF2, by copying the action, correction, and reward patterns that it has stored in memory (e.g., MD3, MD5, and MD4) and assign those patterns to the new ReflexF1 along with the new trigger pattern MD2. Depending on the settings of the floor lamp, the floor lamp may have just learned a new association between the trigger event related to the wall switch toggling on and the activation of the floor lamp's light. For example, the floor lamp may be in a critical learning period 801 (as shown in
When the user turns on the floor lamp shortly after toggling the wall switch, the smart box associated with the floor lamp may correlate that lamp-on event with the new reflex, ReflexF1, with the recently learned pattern, MD2, as its trigger pattern. The actions of the wall switch being turned off and the floor lamp being turned off soon thereafter may generate similar responses in the wall switch and floor lamp.
When the wall switch is toggled a second time from ‘off’ to ‘on’, the associated occurrence data of the on event may again be broadcast in an event report message and received by the floor lamp. Again the floor lamp may process the related event report message with the occurrence data, generating an event and eventually pattern MD2. However, this time the floor lamp matches generated pattern MD2 to the known pattern of ReflexF1. In response to this matching, the floor lamp may also determine that there is a match between pattern MD2 and the stored trigger pattern of ReflexF1 and may enter triggered mode with respect to ReflexF1. Further, the floor lamp may determine whether the trigger weight of ReflexF1 is equal to or above the trigger weight threshold. In this example after only one training cycle, the floor lamp may determine that the trigger weight of ReflexF1 is not equal to or above the trigger weight threshold because ReflexF1 is a new reflex. Thus, the floor lamp may continue to monitor for more events while in the triggered mode for ReflexF1.
When the user turns on the floor lamp within the 5-10 second time window, the floor lamp may generate a reward event that eventually causes the floor lamp to generate reward pattern MD4. The floor lamp processes MD4 and determines that there is a reward pattern match with ReflexF1. In response, the floor lamp, still in the triggered mode with respect to ReflexF1, may increase the trigger weight of ReflexF1. After adjusting the weights or after the 5-10 second time window, the floor lamp may exit the triggered mode with respect to ReflexF1 and enter a monitoring mode where the floor lamp monitors for more events.
Sometime later, the user may toggle the wall switch to on a third time again causing the wall switch to broadcast an event report message with occurrence data that is received by the floor lamp. Again, based on data in the received message, the floor lamp may generate a related event and then pattern MD2. The floor lamp may determine that there is a trigger match between generated pattern MD2 and the trigger pattern of ReflexF1 and may enter the triggered mode with respect to ReflexF1 for a third time. Once again, the floor lamp may determine whether the trigger weight of ReflexF1 is equal to or above the trigger weight threshold. For a third time, the floor lamp may determine that the trigger weight of ReflexF1 is not equal to or above the trigger weight threshold because ReflexF1 is a new reflex. Thus, the floor lamp continues to monitor for more events while in the triggered mode.
Still within the 5-10 second time window of the recent triggered mode of ReflexF1, the user may turn on the floor lamp for third time. In response the floor lamp generates the reward event that eventually causes the floor lamp to generate reward pattern MD4. The floor lamp may process pattern MD4 and determine that there is a reward pattern match with ReflexF1. The floor lamp still in the triggered mode with respect to ReflexF1 may increase the trigger weight of ReflexF1 above the trigger threshold.
Sometime later, when the user toggles the wall switch from ‘off’ to ‘on’ for a fourth time the same sequence of events occurs, only this time the floor lamp may determine that the trigger weight of ReflexF1 is equal to or above its threshold and therefore it generates the action pattern MD3. In response to the action pattern MD3, the floor lamp may generate an associated action event that energizes a motor controller that turns on the light of the floor lamp. Thereafter, the floor lamp will be turned on in response to the user toggling the wall switch from ‘off’ to ‘on’.
When not within proximity of the learning modifier device 150 (i.e., when outside of the broadcast range 1302), the smart boxes 103a-103b may be configured to operate in “default learning states”. In other words, the smart boxes 103a-103b may not utilize modified learning capabilities (e.g., increased/decreased learning rates for calculating triggering weights for reflexes, etc.). Default learning states may be different for individual learning devices. Thus, the default learning state for the first smart box 103a may be defined by a learning rate that is faster than, slower than, or the same as the default learning state for the second smart box 103b. Further, the default learning states may indicate whether the smart boxes 103a-103b are operating within an enabled or disabled learning mode. For example, by default, the first smart box 103a may be configured to operate in a disabled learning mode such that the smart box 103a does not normally generate new trigger-action associations as described above.
Conversely, the smart boxes 103a-103b may be configured to operate in “modified learning states” when they are within reception range of the learning modifier device 150 (i.e., when within the broadcast range 1302). In other words, based on receiving learning modifier signals from the learning modifier device 150, the smart boxes 103a-103b may utilize modified learning capabilities (e.g., adjusted learning rates for calculating triggering weights for reflexes, etc.). For example, the modified learning states may correspond to the smart boxes 103a-103b being configured to learn to perform actions (e.g., increase trigger weights above thresholds, etc.) in response to receiving fewer triggers when in the modified learning states. In some embodiments, the modified learning states may be relative to and thus different for individual learning devices. For example, the modified learning state for the first smart box 103a may be defined by a faster, slower, or the same learning rate as the modified learning state for the second smart box 103b. However, in some embodiments, the modified learning states may be the same for all learning devices. For example, the modified learning state for the smart boxes 103a-103b may be defined as a universal learning rate to cause congruent learning behaviors in both smart boxes 103a-103b. Similar to as described above with reference to default learning states, the modified learning states may indicate whether the smart boxes 103a-103b are operating within an enabled or disabled learning mode. For example, when the first smart box 103a is configured to operate in an enabled learning mode by default, the modified learning state of the first smart box 103a may be a disabled learning mode.
In
In
In
The learning rate modifier value 1608 may indicate whether learning devices receiving the learning modifier signal should increase or decrease their rate of learning. In some embodiments, the learning rate modifier value 1608 may be used as a value in trigger weight calculations as described above, such as a scalar or a multiplier, etc. For example, when a value greater than ‘1’, the learning rate modifier value 1608 may be a scalar applied to trigger weights of reflexes during calculations in order to increase learning rates (i.e., trigger weights are more rapidly adjusted). As another example, when a value less than ‘1’, the learning rate modifier value 1608 may be a scalar applied to trigger weights of reflexes during calculations in order to decrease learning rates (i.e., trigger weights are more slowly adjusted). In some embodiments, the learning rate modifier value 1608 may be an additive value used in trigger weight calculations as described above. For example, the learning rate modifier value 1608 may be added to (or subtracted from) various variables (e.g., gains, bias, etc.) used in calculating trigger weights for reflexes.
The data structure 1600 may also include various optional components 1610-1616. In particular, the data structure 1600 may include an optional learning mode active setting component 1610 that may indicate whether devices receiving the data structure 1600 within a signal should be placed in an enabled learning mode (i.e., capable of learning) or in a disabled learning mode (i.e., incapable of learning). In some embodiments, the learning rate modifier value 1608 may be configured to function in the same manner as the learning mode active setting 1610. For example, a zero value for a learning rate modifier value 1608 may be used by a learning device as a multiplier that zeroes any learning values or weights calculated by the learning device. As another example, a value of ‘1’ for the learning rate modifier value 1608 may be used by a learning device as a multiplier that merely maintains any normal learning values or weights calculated by the learning device.
The data structure 1600 may include an optional device-type component 1612 that may indicate the class or type of learning device that may be affected by the learning modifier signal. For example, the device-type component 1612 may indicate that any or all of smart TVs, smart stereos, smart wall switches, and/or smart lamps may adjust their individual learning modes and/or learning rates in response to receiving a learning modifier signal. In some embodiments, the device-type component 1612 may indicate that all or none of the devices receiving the learning modifier signal may be affected, such as when the corresponding learning modifier device is configured to indiscriminately affect the learning of nearby devices. In some embodiments, the device-type component 1612 may indicate particular device identifiers that may be affected by learning modifier signals including the data structure 1600. For example, the device-type component 1612 may include a range of device identifiers or machine access control (MAC) addresses that may be configured to utilize the information within associated learning modifier signals.
The data structure 1600 may also include an optional learning rate modifier-type component 1614 that may indicate particular types of learning calculations that may be affected by learning modifier signals. The learning rate modifier type component 1614 may indicate that only calculations related to reward patterns and/or correction patterns may be affected by the learning rate modifier value 1608. For example, the learning rate modifier type 1614 may indicate that trigger weights for various reflexes may only be changed at an adjusted rate in response to receiving reward patterns during a triggered mode. As another example, the learning rate modifier type component 1614 may indicate that trigger weights for various reflexes may only be changed at an adjusted rate in response to receiving correction patterns during the triggered mode. In this way, learning modifier signals may be transmitted to modify the rate for learning or unlearning to perform actions in response to triggers.
The data structure 1600 may also include an optional transmission frequency component 1616 that indicates how often the learning modifier device sends the learning modifier signals. Such frequency information may be used by learning devices to determine whether they are no longer receiving learning modifier signals (e.g., out of range of the learning modifier device, etc.) or are merely within a period between signals. For example, a smart TV may determine that it may no longer be within a modified learning state when a learning modifier signal has not been received within a time window defined by a transmission frequency component 1616 from a first learning modifier signal.
In block 1701, the processor of the learning device may monitor a receiver circuit for incoming signals, such as by monitoring an incoming message buffer to detect new short-range wireless messages (e.g., Bluetooth packet) received from nearby devices. At this time, the learning device may not utilize modified learning capabilities, and thus be considered to be operating in a default learning state. In determination block 1702, the processor of the learning device may determine whether a received signal is a learning modifier signal. For example, the learning device may use metadata, header information, and/or any other data in the received signal that may indicate whether the signal was transmitted by a learning modifier device or otherwise includes information that may change the way the learning device is configured to learn. In response to determining that the received signal is not a learning modifier signal (i.e., determination block 1702=“No”), the learning device may continue monitoring for incoming signals to process with the operations in block 1701.
However, in response to determining that the received signal is a learning modifier signal (i.e., determination block 1702=“Yes”), the processor of the learning device may modify one or more of the learning capabilities of the learning device to operate in a modified learning state based on the received signal in block 1704. The learning device may change various settings, configurations, values, and other data stored on the device in order to adjust learning behaviors while in the modified learning state. Specific examples of modified learning capabilities are described below with reference to
In determination block 1706, the processor of the learning device may determine whether subsequent learning modifier signals are received. For example, the learning device may monitor a receiver circuit for incoming signals that are received subsequent to the time the learning device modified its learning capabilities (i.e., entered a modified learning state) based on the received learning modifier signals. The learning device may determine whether subsequent learning modifier signals are received within a certain time frequency and/or within a certain time window from the receipt of a preceding learning modifier signal. In other words, subsequent learning modifier signals may need to be received close enough together to maintain the learning device operating in the modified learning state. For example, the learning device may remain configured to operate in a modified learning state when a second learning modifier signal is received within a predefined time period (e.g., a second, etc.) after receiving a first learning modifier signal, a third learning modifier signal is received within the time period after receiving the second learning modifier signals, and so on. In some embodiments as described below with reference to
In response to determining that subsequent learning modifier signals are received (i.e., determination block 1706=“Yes”), the learning device may remain in the modified learning state with the modified learning capabilities and may continue with the operations in determination block 1706. However, in response to determining that subsequent learning modifier signals are not received or are not received within particular time periods after the first learning modifier signal was received (i.e., determination block 1706=“No”), the processor of the learning device may reset the learning capabilities of the learning device to operate in the default learning state in block 1708. In other words, if there have been any adjustments to information (e.g., variables, configurations, modes, settings, etc.) associated with the calculations, routines, configurations, and/or other functionalities for creating associations between triggers and actions (i.e., generating reflexes) and/or changing trigger weights of learned associations (i.e., increasing or decreasing trigger weights of stored reflexes), the learning device may return the adjusted information to its default settings, condition, content, and/or values. For example, the learning device may negate, remove, zero-out, or otherwise reset multipliers received in learning modifier signals that were used to adjust calculations of trigger weights.
As an illustration, a learning device may start a timer in response to modifying learning capabilities (e.g., increased learning rate, etc.) based on a first learning modifier signal received from a nearby learning modifier device. Before the timer elapses, the learning device may receive a second learning modifier signal and may reset the timer. When a user physically moves the learning modifier device away from the learning device beyond the broadcast range of the learning modifier device (e.g., into another room or building), the learning device may will not receive subsequent learning modifier signals, and eventually the timer will elapse. As a result, the learning device may discontinue using modified learning capabilities and return to a default learning state (e.g., default learning rate, etc.).
In some embodiments, the timer may be used to measure a predetermined time period, such as a standard period for all learning devices. In some embodiments, the time period may be different for different learning device types (e.g., smart TV, smart stereo, etc.), manufacturers, and/or learning modifier devices. In some embodiments, the time period may be based on the reliability of a communication channel and/or communication protocol used for learning modifier signals, such as a first time period for learning modifier signals transmitted via Bluetooth and a second time period for learning modifier signals transmitted via WiFi, etc. In some embodiments, the time period of a timer may be based on information within signals received at a learning device. For example, the default (or maximum) value of a timer (i.e., the timer period represented by the timer) may be set based on data from a learning modifier signal that indicates a broadcast frequency of a nearby learning modifier device.
In block 1751, the processor of the learning device may initialize a timer when the learning device begins operating in a modified learning state. The timer may initially be set to have an expired or otherwise inactive setting, thus indicating the learning device is operating in its default learning state and not operating in a modified learning state. As described below, the timer may be activated in response to receiving learning modifier signals from a learning modifier device. Once activated and not expired, the timer may be regularly adjusted (e.g., incremented, decremented, etc.) by the learning device. In some embodiments, the timer may be configured to increase in value up to a maximum value or alternatively may be configured to decrease in value down to a minimum value. For example, the learning device may increment (i.e., count up) or decrement (i.e., count down) the timer's value based on a clock signal, every millisecond, every second, etc.
In some embodiments, the learning device may utilize a plurality of timers, each associated with a different learning modification. For example, the learning device may use a first timer to indicate whether the learning device is operating in a modified learning state related to a first reflex and a second timer to indicate whether the learning device is operating in a modified learning state related to a second reflex.
In determination block 1752, the processor of the learning device may determine whether the timer has elapsed. For example, when the timer is configured to count down to a minimum value, the learning device may determine that the timer has elapsed when the current value of the timer is equal to or below the minimum value (e.g., zero). As another example, when the timer is configured to count up to a maximum value, the learning device may determine the timer has elapsed when the current value of the timer is equal to or above the maximum value. Until a first learning modifier signal is received and the learning device has entered the modified learning state, the learning device may always determine the timer has expired as it has yet to be activated.
In response to determining that the timer has elapsed (i.e., determination block 1752=“Yes”), the processor of the learning device may reset the learning capabilities of the learning device to operate in a default learning state in block 1708 as described above. In some embodiments, the operations in block 1708 may only be performed when the learning device has modified learning capabilities. In other words, the operations in block 1708 may be superfluous when the timer has not been activated and therefore may be skipped. For example, when the learning device has already performed the operations in block 1708 in a previous iteration of the operational loop of the method 1750 and has not subsequently received any learning modifier signals (i.e., has not entered the modified learning state), the values of the variables used to calculated trigger weights may already be set to their default values and thus may not need to be reset again.
In response to determining that the timer has not elapsed (i.e., determination block 1752=“No”), or the operations in block 1708 have been performed, the processor of the learning device may determine whether a signal is received in determination block 1754. For example, the learning device may monitor an incoming message buffer to detect whether a new short-range wireless message (e.g., Bluetooth packet) from a nearby device has been received. In response to determining that a signal has not been received (i.e., determination block 1754=“No”), the learning modifier device may continue with the operations for evaluating the timer in determination block 1752.
In response to determining that a signal has been received (i.e., determination block 1754=“Yes”), the processor of the learning device may evaluate the received signal to identify a source and/or a message type of the received signal in block 1756. In particular, the learning device may evaluate the data represented within the received signal to identify the various data elements described above with reference to
As described above, in determination block 1702 the processor of the learning device may determine whether the received signal is a learning modifier signal. In particular, the learning device may make this determination based on the evaluation of the received signal (e.g., the identified source and/or message type) as described above with reference to the operations in block 1756. For example, the learning device may determine that the received signal is a learning modifier signal in response to matching a device identifier from the received signal to a stored identifier associated with an approved learning modifier device. As another example, the learning device may determine that the received signal is a learning modifier signal based on identifying a code or descriptor within the received signal that corresponds to a learning modifier signal (e.g., a signal type code pre-associated with learning modifier devices, etc.). In some embodiments, the learning device may determine that the received signal is a learning modifier signal in response to identifying commands, codes, information, etc. that indicate a learning modification, such as the presence of multiplier values for gains. For example, a signal may be determined to be a learning modifier signal when including a script or command that relates to disabling or enabling a learning mode and/or changing values of variables used in calculating trigger weights of reflexes stored on the learning device.
In response to determining that the received signal is not a learning modifier signal (i.e., determination block 1702=“No”), the processor of the learning device may process the received signal as an event report message in block 1758. In other words, the learning modifier device may handling the received signal by performing the operations in the method 1100 of
In response to determining that the received signal is a learning modifier signal (i.e., determination block 1702=“Yes”), the processor of the learning device may determine whether the received learning modifier signal is applicable to the learning device in optional determination block 1760. The determination in optional determination block 1760 may be optional as in many cases a learning modifier signal may cause any learning device to enter a modified learning state. However, depending on the information included with the received learning modifier signal, the learning device may determine the received learning modifier signal is not relevant. For example, based on device types (e.g., smart TVs, smart stereos, etc.) indicated in the received learning modifier signal, the learning device may determine that the signal is not applicable as the learning device is not one of the indicated device types (e.g., the learning device may be a smart lamp, etc.). As another example, based on time or formatting information within the received learning modifier signal, the learning device may determine that the signal is not applicable as the time is not within the learning device's time window for evaluating signals and/or the format is not compatible with the learning device. As another example, the received learning modifier signal may not be applicable when it includes a secret code or key (e.g., a trusted code) that does not match data stored on the learning device.
In some embodiments, the learning device may determine that the received learning modifier signal is not applicable when the device is incapable of entering a modified learning state. For example, a learning device with a physical setting (e.g., toggle, switch, lever, etc.) associated with the learning device's learning mode may be incapable of entering a modified learning state when that physical setting is set to ‘off.’
In response to determining that the received learning modifier signal is not applicable to the learning device (i.e., optional determination block 1760=“No”), the learning modifier device may continue with the operations for evaluating the timer in determination block 1702. In response to determining that the received learning modifier signal is applicable to the learning device (i.e., optional determination block 1760=“Yes”), the processor of the learning device may modify one or more of the learning capabilities of the learning device to operate in a modified learning state in block 1704 based on the received signal as described above.
In block 1762, the processor of the learning device may activate or reset the timer for the modified learning state. For example, the learning device may activate a timer mechanism configured to continually count down by setting a current timer value to a maximum value. When the timer has already been activated based on previous iterations of the method 1750 and has not expired, the learning device may reset the current value of the timer to its maximum value or minimum value, depending on whether the timer is configured to count down or count up over time. For example, when the timer is configured to continually count down, is activated, and has not expired (e.g., the timer value has not reached a zero value, etc.), the learning device may reset the timer's current value to a maximum value. As another example, when the timer is configured to continually count up, is activated, and has not expired (e.g., the timer value has not reached a maximum or ceiling value, etc.), the learning device may reset the timer's current value to its lowest value (e.g., zero). The learning modifier device may continue with the operations for evaluating the timer in determination block 1702.
In some embodiments, the learning device may set the parameters for the timer based on the data of the received learning modifier signal. In particular, the learning device may store the maximum and minimum values of the timer based on transmission frequency data included within the received learning modifier signals. In other words, the learning modifier device that transmitted the received learning modifier signal may teach the learning device when to expect the next learning modifier signal. The maximum timer value may be set to the transmission frequency, and the minimum timer value may be a zero value. For example, when the transmission frequency indicated in a first learning modifier signal is a number of seconds, the learning device may set the maximum timer value as the number of seconds so that the timer may elapse if the learning device does not receive a second learning modifier signal within that number of seconds from receiving the first learning modifier signal, the learning device may return to a default learning state.
In some embodiments, the learning mode may also control whether the learning device may adjust trigger weights of reflexes. For example, when operating in an active or enabled learning mode, the learning device may increase the trigger weight of a reflex in response to receiving related reward patterns and/or decrease the trigger weight in response to receiving related correction patterns. However, when the learning mode is disabled or deactivated, the learning device may not be capable of altering the trigger weight of the reflex, regardless of receiving reward patterns or correction patterns. In some embodiments, the setting of the learning mode (e.g., enabled or disabled) may indicate both whether the learning device is capable of creating new reflexes and whether the learning device is capable of adjusting trigger weights of existing reflexes.
In
In response to determining that the received learning modifier signal is applicable to the learning device (i.e., optional determination block 1760=“Yes”), the learning device may enable the learning mode, such as by setting a system variable, flag, or other information stored on the learning device in block 1804. For example, the learning device may set a flag that indicates the learning device is in a modified learning state and thus new reflexes may be generated and/or trigger weights of stored reflexes may be adjusted. In response to enabling the learning mode, the method 1800 may continue with the operations in blocks 1762 for activating or resetting the timer for the modified learning state.
The method 1850 illustrated in
In method 1850, in response to determining that the timer has elapsed (i.e., determination block 1752=“Yes”), the processor of the learning device may enable its learning mode in block 1804, and may perform the operations in determination block 1754. In response to determining that the received learning modifier signal is applicable to the learning device (i.e., optional determination block 1760=“Yes”), the learning device may disable the learning mode in block 1802, and may activate or reset the timer for the modified learning state in block 1762.
In response to determining that the received learning modifier signal is applicable to the learning device (i.e., optional determination block 1760=“Yes”), the learning device may enter a modified learning state. Thus in block 1904, the processor of the learning device may adjust the values of variables used to calculate trigger weights of reflexes based on the received signal. For example, based on data indicating one of a learning rate modifier value (e.g., a positive multiplier, a negative multiplier, an adder, etc.), the learning device may add variables to equations for calculating trigger weights and/or change values of pre-existing variables of the equation in order to cause the learning device to calculate the trigger weights in different ways. Such changes to the way the learning device calculates trigger weights for various reflexes may alter the learning rate of the various reflexes. For example, increasing certain values may increase the speed at which the learning device learns to perform an action associated with a trigger. As another example, decreasing other values may slow the speed at which the learning device learns to perform the action associated with the trigger.
In some embodiments, based on the received learning modifier signal, the learning device may make adjustments such that only certain types of calculations may be adjusted. For example, the received learning modifier signal may include a learning rate modifier type that indicates only reward signals and correction signals are to be employed differently, causing the learning device to adjust calculations for rewarding or correcting trigger weights. In response to adjusting the values of the variables, the method 1900 may continue with the operations in blocks 1762 for activating or resetting the timer for the modified learning state.
The foregoing method descriptions and the process flow diagrams are provided merely as illustrative examples and are not intended to require or imply that the steps of the various aspects must be performed in the order presented. As will be appreciated by one of skill in the art the order of steps in the foregoing aspects may be performed in any order. Words such as “thereafter,” “then,” “next,” etc. are not intended to limit the order of the steps; these words are simply used to guide the reader through the description of the methods. Further, any reference to claim elements in the singular, for example, using the articles “a,” “an” or “the” is not to be construed as limiting the element to the singular.
The various illustrative logical blocks, modules, circuits, and algorithm steps described in connection with the aspects disclosed herein may be implemented as electronic hardware, computer software, or combinations of both. To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, modules, circuits, and steps have been described above generally in terms of their functionality. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the overall system. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention.
The hardware, such as smart box 103, is used to implement the various illustrative logics, logical blocks, modules, and circuits described in connection with the aspects disclosed herein may be implemented or performed with a general purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. A general-purpose processor may be a multiprocessor, but, in the alternative, the processor may be any conventional processor, controller, microcontroller, or state machine. A processor may also be implemented as a combination of computing devices, e.g., a combination of a DSP and a multiprocessor, a plurality of multiprocessors, one or more multiprocessors in conjunction with a DSP core, or any other such configuration. Alternatively, some steps or methods may be performed by circuitry that is specific to a given function.
In one or more exemplary aspects, the functions described may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software, the functions may be stored as one or more instructions or code on a non-transitory computer readable storage medium, non-transitory computer-readable medium or non-transitory processor-readable storage medium. The steps of a method or algorithm disclosed herein may be embodied in a processor-executable software module, which may reside on a non-transitory computer-readable or processor-readable storage medium. Non-transitory computer-readable or processor-readable storage media may be any storage media that may be accessed by a computer or a processor. By way of example but not limitation, such non-transitory computer-readable or processor-readable media may include RAM, ROM, EEPROM, FLASH memory, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium that may be used to store desired program code in the form of instructions or data structures and that may be accessed by a computer. Disk and disc, as used herein, includes compact disc (CD), laser disc, optical disc, digital versatile disc (DVD), floppy disk, and blu-ray disc where disks usually reproduce data magnetically, while discs reproduce data optically with lasers. Combinations of the above are also included within the scope of non-transitory computer-readable and processor-readable media. Additionally, the operations of a method or algorithm may reside as one or any combination or set of codes and/or instructions on a non-transitory processor-readable medium and/or computer-readable medium, which may be incorporated into a computer program product.
The preceding description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the following claims and the principles and novel features disclosed herein.
The present application claims the benefit of priority to U.S. Provisional Application No. 61/827,141, entitled “A Method and Apparatus for Continuous Configuration of a Device” filed May 24, 2013, the entire contents of which are hereby incorporated by reference for all purposes. The present application is also related to U.S. patent application Ser. No. 14/286,244 entitled “Learning Device With Continuous Configuration Capability”, which is filed concurrently herewith, the entire contents of which are incorporated by reference for further details regarding learning devices.
| Number | Date | Country | |
|---|---|---|---|
| 61827141 | May 2013 | US |
