Patent Application
20100030711
The present invention relates generally to analog computing and to hybrid digital/analog computing, and more particularly, to generating overlays for these types of computers.
The term evolutionary computation is used herein to refer to machine learning optimization and classification paradigms that are inspired by evolutionary mechanisms such as biological genetics and natural selection. The evolutionary computation field includes genetic algorithms, particle swarm optimization, evolutionary programming, genetic programming, and evolution strategies.
A swarm is used herein to refer to a population of interacting elements that collaboratively search through a problem space in order to optimize some global objective. Problems involving multiple objectives and/or multiple constraints are also optimized. Interactions between relatively local (topologically) swarm elements are often emphasized. Moreover, a swarm tends to have a general stochastic (or near-chaotic) characteristic that causes swarm elements to move toward a center of mass in the population located on critical dimensions, thus resulting in convergence on an optimum (or multiple optima) for the global objective of the swarm.
A particle swarm, as used herein, is similar to a genetic algorithm (GA) in that the system is initialized with a population of randomized positions in hyperspace that represent potential solutions to an optimization problem. However, each particle of a particle swarm, unlike a GA, is also assigned a randomized velocity. The particles (i.e., potential solutions) are then “flown” through hyperspace based upon their respective velocities in search of an optimum solution (or optimum solutions, in case of multiple optima) to a global objective.
Digital microcontrollers are in widespread use. Approximately 4-5 billion were manufactured in 2005. These microcontrollers, such as PICs and 8051s are self-contained computers that are generally programmed to do one task. Personal computers (PCs) and almost all other computers are digital machines. They do not perform tasks that can be performed by analog computers as quickly, or by using simple analog computing circuitry. In some instances, they are substantially more complex and costly than needed if analog capability were utilized.
As is the case with digital microcontrollers, an analog microcontroller would generally be configured (not really programmed as digital microcontrollers are) to do one task. However, the analog microcontroller can be implemented with no RAM, no ROM, no clock, and no program as used in digital microcontrollers. Because of its simplicity, it is less expensive. Because of its implementation using only analog components including a continuous sheet processor, It is also much faster. (Some digital logic on the input and/or the output may be needed in certain applications to interface the analog microcontroller to digital devices/systems.) The sheet processor with multiple pins can be implemented with a wide variety of materials and configurations, including polymer sheets, conductive foam, and silicon.
These analog microprocessors can also be used in combination with digital processors to realize hybrid machines that have capabilities not currently available, or that can provide currently-available capabilities significantly more rapidly and less expensively. The configuration of these hybrid machines, including which tasks are done by digital and which by analog means, and how those tasks are accomplished in an optimal manner, need to be done using a method and apparatus that are flexible, rapid, inexpensive, and able to respond to highly complex and changing environments.
A need therefore exists for a method and apparatus which evolve the configurations for analog microcontrollers and for digital/analog hybrid computers that utilize continuous computing methodology.
A method is described below that may be used to configure an extended analog computer for use as an application controller. The method includes selecting input pins from among a plurality of pins in a continuous sheet processor, selecting an arrangement of intermediate and output pins from among the remaining pins in the plurality of pins in the continuous sheet processor, applying a pattern data set to the input pins, using an evolutionary algorithm, coupling current sources and sinks to the intermediate and output pins, measuring an error between an output and its expected value, and continuing to select intermediate and output pin arrangements, apply pattern data sets, and measure errors until a configuration threshold is met.
An analog microcontroller may be described as an embodiment of an extended analog computer. An extended analog computer may be comprised of a plurality of electrical current injection and/or voltage application points on a continuous resistive sheet. A continuous resistive sheet may include sheet materials such as, but not limited to, silicon, conductive foam, and conductive polymers. Programming of the extended analog computer may be implemented with an overlay. An overlay may be described as a configuration of injected current sources and/or applied voltage sources, and fuzzy logic or Kirchhoff-Lukasiewicz (K-L) functions coupled to other processing elements.
A hybrid extended analog/digital computer (hEAC) is a combination of digital and analog circuitry. Factors such as device cost, device complexity, application flexibility, and system integration requirements are used to select an optimal configuration for implementing an application. The configuration may be adjusted for changing requirements.
For an analog microcontroller, an exemplary method or process for evolving an overlay utilizes an evolutionary algorithm, such as, but not limited to, particle swarm optimization. As noted above, an overlay is comprised of injected current sources and/or applied voltage sources, and fuzzy logic (FL) and/or Kirchhoff-Lukasiewicz function parameters. Moreover, the exemplary method simplifies the number and/or arrangement of the injected current sources and/or applied voltage sources, and/or the fuzzy logic or K-L functions as it adapts to corresponding parameters meeting certain criteria. The method also removes current sources, voltage sources, and/or functions from the overlay as it adapts to corresponding function parameters satisfying certain criteria.
The configuration or overlay generated by the method may be downloaded from a computer or a digital microcontroller in a manner that is analogous to the downloading of a file containing compiled computer code into a microcontroller. The method may be used to implement an artificial neural network on an analog computer. One aspect of the method is to configure a portion or all of the pins not reserved for inputs or outputs by evolving the current source (+) or sink (−) values and/or voltage sources that are coupled to the pins. For example, in one embodiment having a 25-pin R002 EAC board, input and output pins are selected and the functions and current sources/sinks for the remaining pins are evolved by the method. For example, each pattern in a 150-pattern Iris data set that represents three species of iris flowers may be generated from four variables (inputs). Therefore, four pins on one side of a continuous resistive sheet may be designated as inputs, and either 1 or 3 pins somewhere else on the sheet may be designated as outputs. The functions and/or current sources/sinks for other pins may be evolved. The evolution may also include revising the input/output pin designations. If three outputs are used in the example, the one generating the greatest current or voltage may be used to identify a class, although other ways of identifying a class may be used. If one output is used, fuzzy membership functions may be activated by the output. One example of the embodiment of an analog microcontroller appears in block diagram form in
As shown in
The analog microcontroller 10 may have no digital components, although digital components may be used. The core computing performed by the microcontroller has been estimated to take less than a nanosecond. In analog implementations, the controller is often augmented with fuzzy logic membership functions or Lukasiewicz Logic Arrays. These functions/arrays are used in two primary ways, although other ways are possible. For one, output from an output pin may be passed through one of these functions/arrays. For another, output from an “intermediate” pin, that is, a pin which is neither an input nor an output, may be passed through one of these functions/arrays, and the result re-injected into another intermediate pin, or into an output pin. In addition to systems configured with one analog processor sheet, systems may include arrays of these sheets for significant extensions of the controller's computing capability. These sheets may be arranged in two dimensions, as a flat sheet of multiple processor sheets, or as a vertical column of multiple processor sheets. They may also be arranged in three-dimensional arrays, such as 10×10×10 (or larger) arrays. The various sheets may be located proximally to the controller, or may be geographically distributed in multiple locations and linked by communication links, such as the internet or other communication methodologies. Embodiments described below use evolutionary algorithms to implement a continuous distributed computational model that replaces a sequential digital computational model. This model is realized with structures that are evolved using evolutionary computation paradigms.
A process that may be used to develop an overlay for an extended analog computer (EAC) being used as a controller is now described. The process may be used to develop an overlay for an EAC incorporated into an analog/digital hybrid system, as well as for an EAC used as an analog microcontroller. The process is an example only as other processes may also be used. A controller in a control system is usually configured as operating on a number of input parameters. Different values of these inputs correspond to different output signals. Each time the controller is used, the controller receives a certain combination of inputs and provides, in response, an output, such as a control signal output in a PID-controller or a fuzzy controller.
To use an EAC as a controller, the input data pattern for each input parameter is characterized as an input current or input voltage for the EAC. Each input is provided within a range feasible for the EAC being used. Each parameter may be represented by current injected into one pin (or a voltage applied to one pin), although other representations are possible. The inputs may be passed to output pins via other pins on the EAC. Fuzzy logic or K-L functions may be coupled to input pin(s), intermediate pin(s,) and/or output pin(s).
The entire configuration, including the selection of the input pins, intermediate pins, and output pins, may be evolved. The example of the method now described, however, has the user select the pin configuration. The example process includes:
This process may be used to select a configuration for operating an EAC as a fuzzy controller. For example, an EAC may be used as an elevator brake fuzzy controller. The elevator brake control signal is determined by the speed of the elevator and the distance from the desired stop point. Consequently, two pins on the EAC were selected as inputs and a third pin was selected as the output. Four other pins were selected to be intermediate current sources. The above paradigm was used to train the EAC to implement a fuzzy control function by adjusting the intermediate current sources. The training process used nine patterns for the two inputs and one output. The two inputs were combinations of the low, medium, and high input values. The output was the fuzzy logic result of the input combination. They are listed in the following table:
The procedure to accomplish this task follows the procedure discussed above. The process includes:
Implementing the above-described method/process enables the realization of an analog microcontroller for a particular application. As already noted, the device may be completely analog or incorporate digital elements. The analog microcontroller may be implemented with no RAM, no ROM, no clock, and no program as required with digital microcontrollers. Rather, an evolutionary algorithm, such as particle swarm optimization, is used to configure the analog microcontroller. The configuration may then be implemented with analog circuit elements and a resistive sheet being used as the processor.
A hybrid EAC (hEAC) is a device that contains a set of contact points arranged in a pattern on a conductive substrate. The contact points may be coupled to functions, such as current injection, current extraction, current measurement, and voltage application and/or measurement. The hEAC also includes a mixture of digital and analog circuitry to configure and operate on the currents/voltages at the contact points. The processing elements in the hEAC act to transform the currents/voltages using a transfer function that may include, but is not limited to fuzzy logic and Lukaseiwicz logic functions. These functions may be implemented in either analog circuitry or synthesized using digital logic.
The split between analog and digital circuitry in the hEAC is a continuum. The system design point for the split between analog and digital circuitry may be designed manually or developed using an evolutionary technique for an optimal solution. Also, although the current embodiments of the hybrid configuration feature separate integrated circuits for the digital and analog processors, both of the functions may reside on one integrated circuit. In a manner analogous to math co-processors being integrated directly onto computer integrated circuits, the analog continuous sheet processor and its accompanying circuitry may be integrated onto the same integrated circuit containing the digital components.
The configuration of the hEAC mirrors the analog microcontroller in the use of evolutionary algorithms. In practice, an hEAC may be used by the evolutionary algorithm to develop the configuration that is instantiated into an analog microcontroller. Additionally, if the flexibility of an hEAC is required for an application, the hEAC may be developed into the device for the application. One example of the embodiment of an analog microcontroller implemented with a hEAC is depicted in block diagram form in
In greater detail, the communication interface 54 couples the microcontroller 58 to a personal computer 80, for example. The communication interface 54 may be implemented with a serial or USB link 84, a serial or USB to TTL converter 88, and/or a universal synchronous/asynchronous receiver/transmitter 90, as shown in the figure. One embodiment of the microcontroller 58 may be a MSP430 processor. The input/output circuitry 60 may include digital signal I/O circuitry 94, analog/digital converters 98, and digital/analog converters 100. The digital I/O circuit is coupled to the digital/analog converters 100, which generate analog signals that are used to operate the current sources/sinks 64. The current sources/sinks 64 are also coupled to pins in the conductive sheet processor 104 and to the FL/LLA functions 78. In this embodiment, the current sources/sinks 64 receive output signals from the FL/LLA functions 78. Other embodiments may be arranged differently, and may include features not shown in the depicted embodiment, such as voltage sources. The current sources and sinks coupled to the sheet processor 104 operate in accordance with these output signals as well as the signals received from the digital/analog converters 100. The digital I/O circuitry is also coupled to the current latches 68. These latches provide currents to the sheet processor and outputs from the sheet processor are coupled to the voltage sensors 70 and the current sensors 74. Some of the voltage signals and the current signals are provided from the voltage sensors and current sensors, respectively, to the analog/digital converters 98 for the conversion of these analog signals to digital values that are read by the controller. Some of the currents from the current sensors may be provided to the FL/LLA functions. As already noted some outputs of the FL/LLA functions are provided to the current sources/sinks. Other outputs may be provided to the analog/digital converters for conversion to digital values read by the controller. The computer 80 is not part of the hEAC, but couples the hEAC to other systems for various purposes.
Again with reference to
For current sourcing/sinking in this embodiment, a Howland topology current supply is coupled to a cell, although other types of current sources may also be used. The Howland type of current supply uses a single operational amplifier to provide positive and negative current to the cell by varying a control voltage. The input of the supply is biased so that voltages above about 1.6V sink current and voltages below about 1.6V source current. A digital-to-analog converter is serially connected to the operational amplifier so the microcontroller provides the control voltage that determines whether a cell sources or sinks current.
The LLA/Fuzzy functions 78 shown in
The process described above may be used to develop an overlay for an extended analog computer (EAC) operating as a classifier. The process may be used for an EAC incorporated into a hybrid system, as well as for an EAC used as an analog microcontroller. The process is exemplary only as other processes may also be used.
A pattern classification problem is typically configured as operating on a number of input parameters with input configurations corresponding to classes. A classifier usually generates outputs to identify one class from a group of possible classes. The output of a classifier may, alternatively, classify an input configuration as being or not being a member of a class. In the classifier described below, the classifier receives a certain configuration of the inputs, and provides as output the classifier's estimate as to which class best represents the input configuration. A single “best fit” class may be indicated, or, the relative fit of more than one, perhaps, even all of the classes, may be indicated in some way, such as a rank ordering of the classes.
To use the EAC as a classifier, the input pattern data for each input parameter is characterized as an input current or input voltage for the EAC. Each input is within a feasible range for the EAC. Each parameter may be represented by current injected into one pin or voltage applied to one pin, although other arrangements may be used. The inputs may be passed to output pins via other pins on the EAC, and fuzzy logic or K-L functions may be used on the input pin(s), intermediate pin(s,) and/or output pin(s).
While the entire configuration, including the determination as to which pins are inputs, intermediate pins, or output pins, may be evolved, the following process has the user select the pin configuration. In this example, the process operates in the following manner:
The above process was implemented to solve the XOR problem. In this case, two pins on the EAC were selected as inputs and a third pin was selected as the output. Four other pins were selected to be intermediate current sources. The process described above was used to train the EAC to implement a XOR function by adjusting the intermediate current sources. The training process involves four patterns for the two inputs and one output. The two inputs were combinations of low and high inputs. The output was the XOR result of the input combination. They are listed in the following table:
The procedure to generate a configuration for this task may include:
After the training was finished, a configuration was selected. The configuration identified the locations of the input, output, and intermediate pins, and the values for the injected current sources. In the given example, only seven pins were used to solve the problem since it was a relatively easy problem. For more complicated problems, more pins may be used and the particle swarm optimization may use all of the pins in the sheet processor to solve the problem.
The continuous portion of the process begins with the random generation of intermediate sources/sinks and logic functions. The sources/sinks that may be coupled to intermediate pins may be current injection sources or sinks or applied voltages. The logic functions that may be coupled to intermediate pins include K-L functions and/or other fuzzy logic functions. The random selections of the sources/sinks and logic functions to be coupled to the intermediate pins may be performed with reference to a white noise generator. Following the coupling of sources/sinks and logic functions to the intermediate pins, the outputs are compared to the instantiated targets (block 210.) Preferably, this comparison is performed using analog comparators. An analog error is computed for each particle (block 214.) The errors are compared and the minimum error is selected (block 218.) Preferably, the comparison of the errors is performed with analog comparators. The minimum error is then compared to the best error for the particle to determine whether it is equal to or less than the best error obtained (block 220.) If the minimum error is greater than the best error obtained, an optimized configuration has not been obtained and the process continues by randomly generating another configuration (block 208.)
If the minimum error is as good as or less than the best error achieved by the process, the best error is updated (block 224) and the best error is compared to an optimization threshold (block 228.) If the best error is less than the threshold, the configuration for the intermediate pins is stored in memory (block 230.) Otherwise, the process continues to randomly generate configurations for the intermediate pins and evaluate the configuration to determine whether the configuration generates an error that is less than the optimal threshold. While the process has been described with reference to a threshold that may be used to terminate the process, the process, alternatively, may store an optimal configuration, and then continue the process to determine whether other optimal configurations exist. The best error associated with a configuration may be used to determine whether an optimized configuration solution has been developed and then stored as an alternative configuration, before continuing the process.
Those skilled in the art will recognize that numerous modifications can be made to the specific implementations described above. While the embodiments above have been described with reference to specific applications, embodiments addressing other applications may be developed without departing from the principles of the invention described above. Therefore, the following claims are not to be limited to the specific embodiments illustrated and described above. The claims, as originally presented and as they may be amended, encompass variations, alternatives, modifications, improvements, equivalents, and substantial equivalents of the embodiments and teachings disclosed herein, including those that are presently unforeseen or unappreciated, and that, for example, may arise from applicants/patentees and others.
This application is filed under 35 U.S.C. 371 claiming priority from international application PCT/US2007/023733, which was filed on Nov. 13, 2007, and which claims priority from U.S. provisional application Ser. No. 60/858,814, which was filed on Nov. 14, 2006.
| Filing Document | Filing Date | Country | Kind | 371c Date |
|---|---|---|---|---|
| PCT/US07/23733 | 11/13/2007 | WO | 00 | 5/14/2009 |
| Number | Date | Country | |
|---|---|---|---|
| 60858814 | Nov 2006 | US |
