The present invention relates to an automated system and method for generating learned models of behavior of an actor and/or the actor's environment. More particularly, it relates to a system and method that generates learned models of behavior relating to particular actor/environment events and activities for use in adapting operations of an automated human actor environment monitoring system that may include response capabilities.
The evolution of technology has given rise to numerous, discrete devices adapted to make actor environment more convenient. For example, in the realm of in-home environments, companies are selling microwaves that connect to the internet, and refrigerators with computer displays, to name but a few. These and other advancements have prompted research into the feasibility of a universal home control system that not only automates operation of various devices or appliances within the home, but also monitors activities of an actor in the home. It could then perform device control based upon the actor's activities and/or events in the living space. In other words, it may now be possible to provide coordinated, situation-aware, universal support to an actor in a daily living environment.
The potential features associated with the “intelligent” home described above are virtually limitless. By the same token, the extensive technology and logic obstacles inherent to many desired features have heretofore prevented their implementation. In general terms, the automated monitoring and response system will consist of a suite of sensors in the living space to detect actor and environment states, a set of actuators that control devices in the environment as well as facilitate communication with the actor (and others), and a computer system (local or remote) capable of making decisions to assist the actor. With these parameters in mind, current response and monitoring systems rely on a pre-programmed computer system that executes predetermined operations based upon a predetermined number and located sensors and actuators within the actor's home, each sensor and actuator having known capabilities. These systems are highly inflexible. From a market viability standpoint, such a configuration is simply impractical. That is to say, virtually every installation of the monitoring and response system will entail a different environment layout, and varying numbers and capabilities of sensors and actuators. An inflexible control system is simply unable to adapt to varying sensor and actuator configurations, and thus is of reduced value. Along these same lines, currently-envisioned automated monitoring and response systems rely upon a highly labor intensive initial configuration, whereby the actor and/or others are required to personally “set-up” the operational parameters (e.g., designate types, locations, and capabilities of sensors and actuators; designate typical or expected daily activities of the actor; designate daily needs of the actor; etc.). Similarly, the capabilities, needs and responses provided by the system will almost certainly need to change over time due to changes in the actor and/or the actor's environment. Currently, inflexible systems require that the actor and/or others not only independently recognize that a change in the actor's needs or activities has occurred, but also that the system be manually re-programmed to appropriately address the change(s). For a “technophobic” actor, the initial configuration is overwhelming at best, as would be any subsequent system reconfiguration to address inevitable changes in the actor or in the actor's environment.
Emerging sensing and automation technology represents an exciting opportunity to develop an independent in-home assistant system. In this regard, a highly desirable attribute associated with such a system is an ability to be rapidly deployable, easy to configure, and automatically adapt as the actor, the actor's environment, and technology changes. Unfortunately, current techniques for configuring an appropriate system are highly inflexible, and difficult to configure. Thus, a need exists for a system and method for learning properties (e.g., patterns, profiles, preferences, etc.) of an actor or their environment (including other humans/animals in the environment) as part of the operation of an automated actor environment monitoring and response system.
The present invention relates to a system and method of generating a learned model of behavior of an actor or the actor's environment as part of the operation of an automated monitoring and response system. In this regard,
The key component associated with the system 20 of the present invention resides in the machine learning module 22 as described below. As such, the sensors 26, the actuators 28, as well as other modules (e.g., the situation assessment module 32 and the response planning module 34) can assume a wide variety of forms. Preferably, the sensors 26 are networked by the controller 24. The sensors 26 can be non-intrusive or intrusive, active or passive, wired or wireless, physiological or physical. In short, the sensors 26 can include any type of sensor that provides information relating to the activities of the actor 36 or other information relating to the actor's environment 38. For example, the sensors 26 can include motion detectors, pressure pads, door latch sensors, panic buttons, toilet-flush sensors, microphones, cameras, fall-sensors, door sensors, heart rate monitor sensors, blood pressure monitor sensors, glucose monitor sensors, moisture sensors, light level sensors, smoke/fire detectors, thermal sensors, water sensors, seismic sensors, etc. In addition, one or more of the sensors 26 can be a sensor or actuator associated with a device or appliance used by the actor 36, such as a stove, oven, television, telephone, security pad, medication dispenser, thermostat, etc., with the sensor or actuator 26 providing data indicating that the device or appliance is being operated by the actor 36 (or someone else).
Similarly, the actuators or effectors 28 can also assume a wide variety of forms. In general terms, the actuators or effectors 28 are configured to control operation of a device in the actor's environment 38 and/or to interact with the actor 36. Examples of applicable actuators or effectors 28 include computers, displays, telephones, pagers, speaker systems, lighting systems, fire sprinkler, door lock devices, pan/tilt/zoom controls on a camera, etc. The actuators or effectors 28 can be placed directly within the actor's environment 38, and/or can be remote from the actor 36, providing information to other persons concerned with the actor's 36 daily activities (e.g., caregiver, family member(s), etc.). For example, the actuator 28 can be a speaker system positioned in the actor's 36 kitchen that audibly provides information to the actor 36. Alternatively, and/or in addition, the actuator 28 can be a computer located at the office of a caregiver for the actor 36 that reports desired information (e.g., a need to refill a particular medication prescription).
The controller 24 is preferably a microprocessor-based device capable of storing and operating the various modules illustrated in
With the above in mind, the machine learning module 22 preferably provides a means for on-going adaptation and improvement of system 20 responsiveness relative to the needs of the actor 36. The machine learning module 22 preferably entails a learned behavior model built over time for the actor 36 and/or the actor's environment 38 (including persons other than the actor 36). In general terms, the learned model is built by accumulating passive (or sensor-supplied) data and/or active (actor and/or caregiver generated) data in an appropriate database. The resulting learned models are then available to the controller 24 (or modules therein) for enabling the system 20 to automatically configure itself and to adapt to changing conditions, minimizing the time and labor involved in set-up and maintenance. Further, the learned models of behavior can be employed to assist in selecting a most appropriate response plan, including the quality and presentation of the interaction. Even further, the learned models of behavior can be utilized by other monitoring and response systems for improved default installation parameters and designations. In short, the learned models of behavior can be used to improve general system performance.
The machine learning module 22 is preferably configured to generate learned models of behavior for a number of different events, activities, or functions (of the actor 36, the environment 38, or the system 20). The goal of the machine learning module 22 is to model the actor's 36 and/or the actor's environment 38 behavior with enough accuracy to recognize situations and respond to situations. The data used by the machine learning module 22 in generating these learned models of behavior is provided by the sensors 26 (unsupervised learning) and/or feedback from the actor 26 or others (supervised learning). To this end, and as used throughout the specification, reference is made to “firing” of a sensor. This is intended to mean that a particular passive sensor has been “activated” or otherwise sensed an action (e.g., a motion sensor “fires” when something moves through, for example, a light beam; a glass break sensor “fires” when the glass to which the sensor is associated with breaks, etc.), or that an active “sensor” has received information indicative of a particular action or activity. With this in mind, examples of learned models of the actor include patterns of the actor 36 over a week (such as waking/sleeping times, toilet use), activity schedules (bathing, cooking, recreation, etc.), paths normally followed by the actor 36 when moving from room to room, preferences for and efficacy of interaction models, etc. Learned models of the environment 38 include, for example, patterns such as expected background noise, patterns of visitors to the home, appliance power use signatures, efficacy of interaction models with other actors, etc. Learned models relating to performance of the system 20 include, for example, sensor reliability and actuator response time (e.g., how long it takes to clear smoke from a particular room). Similarly, the effectiveness of learned models can be learned, as well as a breadth of system coverage (e.g., how much learning is needed a priori or based on learned models from other system installations). Further, the learned models can indicate system sensitivity to changes (e.g., how quickly the system 20 must adapt to changes in an elderly actor) and robustness in the presence of a typical events.
A variety of techniques can be employed by the machine learning module 22 to build the learned models of behavior (e.g., actor's behavior, environment's 38 behavior, or any other “tracked” persons or things in the environment 38). In general terms, the learned models of behavior can be built using calibrated, simple parametric models; trending, averages, and other statistical measures; sequences of or correlations between sensor/actuator firings/usage; sequences or structure of inferred activity models; timings associated with the above and/or probability transitions associated with the above.
Regardless of how the machine learning module 22 generates learned models of behavior, the so-generated information can be utilized in a number of fashions. For example, (1) the system 20 can incorporate several machine learning modules, with each module being adapted to generate data for one or more other modules or agents (i.e., the situation assessment module 32, the response planning module 34, etc.) that is otherwise adapted to utilize the generated information. For example, an alarm module can be provided as part of the system 20 that is adapted to raise an alarm for unexpected activities (e.g., the back door being opened at 3:00 a.m.). In a preferred embodiment, this alarm module functions by understanding when an activity is “normal” based upon the learned models of behavior generated by the machine-learning module 22. Similarly, (2) the learned models of behavior information can be utilized to raise an alarm for expected activity that does not otherwise happen (e.g., the actor 36 normally gets out of bed by 8:00 a.m. as determined by the machine learning module 22, on a particular day when the actor 36 does not get out of bed by that time, an alarm would be raised). Further, (3) the machine learning module 22 can be utilized in generating probabilities for the likelihood that an activity “will be engaged in”, with the information being used by an intent recognition module or agent that would bias its recognition of an activity accordingly (e.g., if the machine learning module 22 determines or indicates that it is extremely likely that a cooking activity happens at 5:00 p.m., then the intent recognition module is more likely to recognize a cooking activity at that time even if the sensors are providing information that does not fully substantiate such a conclusion).
Also, (4) the machine learning module 22 can be utilized in building probabilities that tasks will be abandoned (e.g., the actor 36 is two-thirds finished making breakfast, the phone rings, and the actor 36 then forgets to finish breakfast; the machine learning module 22 builds a model of the normal “breakfast making” steps from which a determination can be made that because one or more remaining, identified steps have not been completed, the breakfast making activity has been abandoned). Additionally, (5) the machine learning module 22 can build probabilities for method choice (e.g., the actor 36 can make breakfast either by using the microwave or the stove or the toaster; what are the probabilities of each of these occurring). Further, (6) the machine learning module 22 can build probabilities of action success (e.g., that the actor 36 will reach a desired person by telephone at a certain time of the day).
Another application is in (7) determining and applying learned models of paths the actor 36 normally follows at certain times in the day, and then be useful for anticipating the actor's 36 path and effectuate turning on relevant lights, opening doors, etc. Similarly, (8) the learned models can determine normal timings between certain, regular activities. For example, the learned models can postulate that the actor 36 normally eats lunch three hours after breakfast. This timing information can then be utilized in anticipating further events (e.g., in furtherance of the previous example, if it is determined that the actor 36 ate breakfast at a time of day later than normal, the timing information can be used to anticipate that lunch will also be consumed at a time later than normal).
Yet another possible application of the learned models of behavior is (9) understanding the effectiveness of a particular plan generated by the response planning module 34. For example, the machine learning module 22 can be utilized to determine the most effective recipient (e.g., the actor 36 or other human) of a particular message; the most effective device on which to deliver a message; the most effective modality for a message; the most effective volume, repetition, and/or duration within a modality; and the actor's 36 preferences regarding modality, intensity, etc. In this regard, the mechanism for learning might be very simple at first, but could become increasingly sophisticated by using contextual conditions for lessons (e.g., “the audio messages are ineffective when the actor is not in the same room as the speaker”).
Yet another possible application of the learned models of behavior is (10) developing/sharing of learned models across two or more separate system installations. It is recognized that a particular learned model will likely be specific to a particular actor/environment. However, a learned model of behavior developed for a first system 20 could be used to improve the default behavior of a second system installation for a different actor. For example, the “normal” behavior of a “normal” actor/environment can be modeled, and then used as a baseline to learn the behavior of a second actor/environment. The learning process associated with the second actor/environment will then be much faster. In short, shared learning across two or more systems enables global lessons to be learned (server side learning). Even further, (11) a meta-model of behavior can be created by merging at least two learned models from the same or different system installations, with the meta-model being used as part of the operations of one or more of the systems (or an entirely separate system) as a baseline or to assess system coverage (e.g., whether all sensors are being understood; how much of understanding any one system 20 has of its corresponding actor's behavior; etc.).
Many of the examples of learning given above can proceed without any intervention from an authority or oracle, a style of learning known as “unsupervised learning”. The path model provided above is an example whereby paths can be learned passively by the system 20 via observation of room occupancy over time. Other sorts of learning that require feedback from the actor 36, the actor's caregiver, and/or other system modules (“supervised learning”) are also optionally employed in accordance with the present invention. Feedback may take the form of explicit preferences (e.g., the actor 36 rank orders some available options). Alternatively, the feedback information can be derived from specific statements provided by the actor 36 (e.g., “I prefer a phone to a pager,” or “Don't remind me to use the bathroom when I have guests”.). The feedback may also involve a teacher labeling training cases for the student module/agent. An example of this is a teacher telling a learning acoustic monitoring module/agent that “the sounds you heard between 8:40 and 8:55 was the actor preparing breakfast in the kitchen”. Alternatively, feedback from a person other than the actor 36 can be response(s) to questions posed by the system 20. Finally, feedback may be non-specific reinforcement indicating relative success or failure to meet a system goal. Reinforcement learning could be particularly applied to the problem of learning the effectiveness of plans or actions. Learned actor preferences can be weighed against measured effectiveness. Using a plan known to be more effective may in some cases trump the desire to conform to the actor's preferences.
Because adaptive systems changes themselves after installation, the present invention preferably includes mechanisms to ensure safety and reliability of the installed system 20. Learned knowledge is checked against built-in knowledge to prevent the system 20 from acquiring dangerous or “superstitious” behaviors. The adaptive components of the system 20 allow the configure, and optionally the actor 36, to disable learning, check point their lessons (freeze), or re-set their learned state to a previous check point, including factory defaults.
One preferred learned model behavior building technique in accordance with the present invention is sequential pattern learning adapted to learn what sensor firings (or other stored information, such as the feedback information described above) correspond to what activities, in what order, and at what time. The technique of sequential pattern mining provides a basis for recognizing these activities and their order. A “sequential pattern” is defined as a list of sensor firings ordered in time. Sequential pattern mining addresses the problem of finding frequent sequential patterns. In a preferred embodiment, the sequential pattern learning associated with the machine learning module 22 extends currently available sequential pattern techniques to incorporate reasoning about the time of the activities.
To better understand the preferred sequential pattern learning technique associated with the machine learning module 22 of the present invention, the foregoing discussion incorporates hypothetical installation details for the system 20, and in particular, types of sensors 26 and environment 38 attributes. These hypotheticals are in no way limiting as the system and method of the present invention can be applied in a wide variety of settings.
One particular hypothetical system 20 installation includes a motion sensor positioned in a bathroom, a motion sensor positioned in a bedroom, a motion sensor positioned in the kitchen, and a motion sensor positioned in a living room. Under this scenario, examples of discovered sequential patterns over a certain number of days include: (1) 75% of the time, the bathroom motion sensor fires after the bedroom motion sensor fires, and (2) a pattern of the kitchen motion sensor firing followed by the living room sensor firing followed by the bedroom motion sensor firing occurs in 60% of the monitored days. From an implementation standpoint, it is not enough to simply apply the sequential pattern algorithm to sensor data and hope to discover sequential patterns that will meaningfully model the actor's 36 behavior and/or behavior of the actor's environment 38. This is because an accurate explanation of what each discovered pattern represents cannot be provided. For example, relative to the above hypothetical data, firing of the bedroom motion sensor may represent several types of activities, such as the actor 36 waking up in the morning, the actor 36 entering the bedroom in the evening to go to sleep, etc. One preferred technique of attaching a meaning to a pattern is to determine the time periods during which each of the monitored sensors are firing. For example, and again relative to the above hypothetical, the examples of discovered sequential patterns can now be designated as: (1) after a bedroom motion sensor firing between 6:45 a.m. and 7:45 a.m., the bathroom motion sensor fires between 7:00 a.m. and 8:00 a.m. 75% of the observed time; and (2) in 60% of the observed days, the pattern kitchen motion sensor fires between 6:00 p.m. and 6:30 p.m. followed by the living room motion sensor firing between 6:20 p.m. and 7:00 p.m. followed by the bedroom motion sensor firing between 9:00 p.m. and 10:00 p.m.
In light of the above, identification of appropriate time periods for each sensor makes it easier to attach a meaning to the pattern learned through unsupervised learning. In the above examples, the first pattern might represent the actor's 36 waking-up routine, while the second pattern might represent the actor's 36 after work/evening routine.
In order to properly identify intervals of occurrence of each event in a sequential pattern as highlighted by the above examples, the machine learning module 22 preferably identifies appropriate time periods or intervals for each sensor firing. In this regard, one approach is to generate the intervals during the sequential patterns' discovery phase. However, the overhead is intractable, since all possible intervals must be considered. Alternatively, and in accordance with one preferred embodiment of the present invention, the time intervals are predetermined and enumerated before sequential pattern learning occurs. It is recognized that this approach has the potential drawback that the quality of the discovered patterns can be affected by the way the system 20 and/or the module 22 defines the intervals. However, to minimize this possibility, the preferred machine learning module 22 does not predefine “fixed-width” time intervals, but instead determines time intervals based upon the distribution of sensor firings during the day. More particularly, the preferred machine learning module 22 estimates a probability density function, for sensor firings over time of day using smoothing techniques. In general terms, the resulting smoothed probability density function histogram reflects a curve of sensor activities over time, with the area under the curve from time t1 to time t2 equating to the probability that a sensor will fire between t1 and t2. An exemplary probability density function histogram for a sensor is provided in
The preferred time interval determination defines time intervals as intervals between successive local minimas of the probability density function. An important concept underlying this approach is that an actor usually performs each daily activity during a certain time period. For example, the actor 36 may typically wake-up between 8:00 a.m. and 9:00 a.m. A sequence of sensor firings with attached or correlated timings (or “timestamps”) of between 8:00 a.m. and 9:00 a.m. would thus represent a “waking-up” activity.
Because some sensors will fire more frequently than others within a particular installation and will likely measure different kinds of activities, the preferred machine-learning module 22 determines time intervals by determining probability density functions for each sensor separately. For example,
Returning to
The above-described algorithm can be expressed as follows:
Then, for each consecutive pair within M′ (∀ pairs (mi, mi+1)), intervali is designated as [x, w] where:
Regardless of how the time intervals are generated, the stored sensor firing/timestamp information is then reviewed and the appropriate time interval is associated with each entry. For example, a system installation may include four sensors A, B, C, and D. Over the course of several days, the machine learning database stores the sensor information, part of which is set forth below in Table 1:
As previously described, the sensor streams are separated (i.e., four separate lists are established, one each for sensors A-D). The probability density function for each list is then generated (i.e., again, four separate probability density functions would be provided). The time intervals are determined for each probability density function as previously described, such that there are a list of time intervals for each of the sensors A-D. With reference to the one specific example provided in Table 1 above, the time intervals are generated as being for sensor A, interval1=[19:21:48, 23:58:09]; for sensor B, time interval1=[18:00:00, 22:03:37]; for sensor C, interval1=[13:23:15, 21:25:18], interval2=[21:40:41, 23:58:09]; and for sensor D, interval1=[21:26:14, 21:46:14]. The sensor firing/timestamp information is then reviewed, and a corresponding, appropriate time interval is “attached” or otherwise associated with each sensor firing/timestamp. With reference again to the above-examples set forth in Table 1, the attached time interval information results in the following time interval correlation as provided Table 2 below:
With the determined time intervals assigned to each sensor firing/timestamp, the machine learning module 22 is preferably adapted to perform sequential pattern discovery on the generated information to produce a learned model of behavior. In a preferred embodiment, the sequential discovery is executed based upon items and sequences. In this regard, each distinct sensor firing/timestamp (and assigned time interval or time partition pair) is defined as an “item”. A “sequence” is an ordered list of items or sequential pattern. A sequence or pattern α can be denoted as (α1→α2→ . . . →αn) where each sequence element αj is an item. An item can occur multiple times in a sequence.
After the time intervals are determined, the list of daily event sequences is translated into a data set suitable for learning sequential patterns. A sensor event maps to an item, and has a unique identifier. Without loss of generality, it can be assumed that no event sequence has more than one event with the same timestamp, and that the events are sorted by the time of occurrences.
The resulting input database stored by the machine learning module 22 consists of a number of these event sequences. The support of a pattern α is the fraction of sequences (or days) in the database that contain that pattern. According to this definition, the support of a pattern α is counted only once per sequence even though it may occur multiple times. Given a user-specified threshold called “minimum support”, it can be determined that a pattern is “frequent” if its support is greater than the minimum support.
Once the list of sensor firings/timestamps is mapped to a database of sequences, the frequent sequential patterns can be computed efficiently using a variety of known sequential pattern discovery algorithms as set forth, for example, in R. Agrawal and R. Srikant, “Mining Sequential Patterns,” Proceedings of the Eleventh International Conference on Data Engineering (ICDE), pp. 3-14, 1995; M. V. Joshi, G. P. Karypis, and V. Kumar, “Universal Formulation of Sequential Patterns”, Technical Report, University of Minnesota, Department of Computer Science, Minneapolis, Minn. 1999; R. Srikant and R. Agrawal, “Mining Sequential Patterns: Generalizations and Performance Improvements”, Proceedings of the Fifth International Conference on Extending Database Technology, pp. 3-17, 1996; the teachings of which are incorporated herein by reference. Other analyses techniques are also available such as Tree Projection Sequential Pattern Algorithm as set forth in V. M. Guralnik, “Mining Interesting Patterns from Sequential Data”, Ph.D. Thesis, University of Minnesota, Minneapolis, Minn., 2001, the teachings of which are incorporated herein by reference.
In more general terms, and again with reference to the previous examples set forth in Tables 1 and 2 above, a sequence is identified as several “items” occurring in a row (i.e., is an ordered list of items). An ordered list of items as part of a sequence requires that the second item must occur after the first item. With reference to the information set forth in Table 2 above, one example sequence consists of: B 21:21:33 [18:00:00, 22:03:37] - - - C 21:25:18 [13:23:15, 21:25:18] - - - A 21:40:51 [19:21:48, 23:58:09]. In one preferred embodiment, identified “frequent” sequences can be refined for purposes of user presentation whereby the interval start and end times are ordered. With reference to the previously-provided exemplary sequence, the time interval refinement changes the identified sequence to be: B 21:21:33 [18:00:00, 21:25:18] - - - C 21:25:18 [18:00:00, 21:25:18] - - - A 21:40:51 [19:21:48, 23:58:09]. Notably, this refinement step does not affect the accuracy of the algorithm, and is performed strictly to enhance presentation/review of the information to/for human reviewers. Relative to an automated monitoring and response system, however, the refinement described above is unnecessary.
Regardless of the exact technique or algorithm(s) utilized to analyze the sensor data, it is recognized that an enormous number of highly frequent, but not necessarily useful, sequential patterns will be generated. In fact, in data sets with characteristics likely encountered with the monitoring and response system 20, the number of patterns generated by the pattern analysis portion of the machine learning module 22 can be the same magnitude as the number of sensor firings in the data set. A list of several thousand patterns is only marginally more useful than the raw data itself. Thus, in a preferred embodiment, the machine learning module 22 is adapted to find “interesting” sequences that will help to create a model of the actor's 36 behavior or activities during the day.
With the above in mind, the machine learning module 22 preferably further includes filter routines for reducing the number of identified patterns into more meaningful or interesting information. A first preferred filter routine entails a domain heuristic that eliminates patterns that are not close in time, and is thus based on the assumption that the closeness in time of occurrence of events provides evidence in support of those firings being related to one activity. The closer the events occur in time, the more likely that the events belong to the same activity. To this end, it is noted that domain knowledge is fairly fuzzy, such that is difficult to define specific parameter values for “closeness in time”.
In one preferred embodiment, the domain heuristic filter routine provides that a pattern represents two different activities if it contains events that have non-intersecting time intervals. By way of example, the pattern analysis may identify the following patterns:
While the first two patterns represent one activity each (waking up and having lunch, respectively), the third represents a combination of the parts of those two activities. Application of the domain heuristic filter routine generally limits the learned patterns to ones that represent one activity. It should be noted that any patterns “discarded” by this filter can be analyzed by a different technique; for example, to calculate that “lunch” occurs a certain timelag after “breakfast”.
A second preferred filter routine eliminates redundant patterns by removing all sub-patterns that have the same supporting days. The closed set reduces the set of all frequent patterns by several orders of magnitude, while not losing any information. For example, if a pattern A→B→C→D is identified as occurring on days 1, 7, 14, 21, as well as patterns A→B→C and B→C→D as also occurring on days 1, 7, 14, and 21, the two sub-patterns can be eliminated. Notably, and with respect to this same example, if the pattern A→B→C occurs on days 1, 2, 7, 8, 14, 21, 24, it is preferably not deemed redundant by the redundant filter routine, and thus would not be eliminated.
A third filter routine preferably provided by the machine learning module 22 is a domain heuristic that eliminates repetitive sensor firing information. In general terms, this repetitive sensor filter routine exploits the domain knowledge that some sensors will continually fire even though the same event triggered them. For example, if the actor 36 props open the front door during the summer, an associated door-contact switch/sensor will continually fire at regular intervals. To condense stored data, these multiple sequential firings of the same sensor are reduced to a single firing.
The machine learning module 22 can further incorporate additional filtering-type operations to facilitate a better understanding of identified patterns. For example, a clustering technique can be performed on the identified patterns, with each cluster representing different variations of the same activity. For example, M. Steinbach, G. Karypis, and V. Kumar, “A Comparison of Document Clustering Techniques”, K.D.D. Workshop on Text Mining, 2000 provides a vector-spaced k-means algorithm, the teachings of which are incorporated herein by reference. To apply this algorithm with the machine learning module 22, the patterns are mapped to a vector space with the dimensions being distinct event types of all the patterns. A pattern is represented as an n-dimensional vector of zeros and ones, with ones corresponding to all of the events present in that particular pattern. This clustering can be done apart from the pattern learning, and can be based upon numerous other parameters, such as proximity (e.g., location of sensor(s) to one another, time, etc.).
The present invention provides a marked improvement over previous designs. In particular, by providing an automated actor monitoring and response system incorporating at least one machine learning module, the resulting system is readily able to adapt to its environment over time. In one preferred embodiment, a sequential pattern learning approach is implemented within the learning module to generate learned models of behavior.
Although the present invention has been described with reference to preferred embodiments, workers skilled in the art will recognize that changes can be made in form and detail without departing from the spirit and scope of the present invention.
This application is related to, and is entitled to the benefit of, U.S. Provisional Patent Application Ser. No. 60/351,300, filed Jan. 22, 2002; U.S. Provisional Patent Application Ser. No. 60/368,307, filed Mar. 28, 2002; U.S. Provisional Patent Application Ser. No. 60/384,899, filed May 30, 2002; U.S. Provisional Patent Application Ser. No. 60/384,519, filed May 29, 2002; U.S. Non-provisional patent application Ser. No. 10/286,398, filed on Nov. 1, 2002; U.S. Provisional Patent Application Ser. No. 60/424,257, filed on Nov. 6, 2002; a non-provisional patent application filed on even date herewith, entitled “System and Method for Automated Monitoring, Recognizing, Supporting, and Responding to The Behavior of an Actor”, having Ser. No. 10/341,335; and a provisional patent application filed on even date herewith, entitled “System and Method for Automatically Generating an Alert Message with Supplemental Information”, having Ser. No. 60/439,346; the teachings of all of which are incorporated herein by reference.
Number | Name | Date | Kind |
---|---|---|---|
4259548 | Fahey et al. | Mar 1981 | A |
4803625 | Fu et al. | Feb 1989 | A |
4843377 | Fuller et al. | Jun 1989 | A |
4952928 | Carroll et al. | Aug 1990 | A |
5086385 | Launey et al. | Feb 1992 | A |
5228449 | Christ et al. | Jul 1993 | A |
5400246 | Wilson et al. | Mar 1995 | A |
5410471 | Alyfuku et al. | Apr 1995 | A |
5441047 | David et al. | Aug 1995 | A |
6130620 | Pinnow et al. | Oct 2000 | A |
6212510 | Brand | Apr 2001 | B1 |
6437696 | Lemelson et al. | Aug 2002 | B1 |
6549915 | Abbott et al. | Apr 2003 | B2 |
6581037 | Pak | Jun 2003 | B1 |
6582380 | Kazlausky et al. | Jun 2003 | B2 |
6611206 | Eshelman et al. | Aug 2003 | B2 |
6678413 | Liang et al. | Jan 2004 | B1 |
6796799 | Yoshiike et al. | Sep 2004 | B1 |
6952662 | Wegerich et al. | Oct 2005 | B2 |
20020084902 | Zadrozny et al. | Jul 2002 | A1 |
20030004652 | Brunner et al. | Jan 2003 | A1 |
20030058339 | Trajkovic et al. | Mar 2003 | A1 |
20030059081 | Trajkovic | Mar 2003 | A1 |
20040018477 | Olsen | Jan 2004 | A1 |
Number | Date | Country |
---|---|---|
0 558 975 | Feb 1993 | EP |
WO 0175653 | Oct 2001 | WO |
Number | Date | Country | |
---|---|---|---|
20030229471 A1 | Dec 2003 | US |
Number | Date | Country | |
---|---|---|---|
60424257 | Nov 2002 | US | |
60384899 | May 2002 | US | |
60384519 | May 2002 | US | |
60368307 | Mar 2002 | US | |
60351300 | Jan 2002 | US |