JPH0553812A - Information processing system - Google Patents

Abstract

PURPOSE:To obtain an information processing system which effectively integrates a knowledge processor and a neural circuit network and does not require the change of a program corresponding to an application. CONSTITUTION:This system is equipped with a working memory 132 enabling the recording of a symbol and the reference/deletion of this recorded symbol, production memory 133 to record a rule for reloading the symbol recorded in the working memory 132, inference engine 132 to reload the recording of the working memory 132 while referring to the recording of the working memory 132 and the rule of the production memory 133, analyzers 11 and 12 to convent an input signal into a symbol by using neural circuit networks N1 and N2 and to record the symbol in the working memory 132 while integrating these neural circuit networks N1 and N2, control executing device 14 to output signals while referring to the recording of the working memory 132, and inference control neural circuit network 16 to control the inference engine 131.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an information processing system, and more particularly to an information processing system which is constructed by integrating a neural network and a knowledge processing device.

[0002]

2. Description of the Related Art Generally, since a knowledge processing device represented by an expert system generally represents knowledge by a symbol, in the case of configuring an information processing system using such a knowledge processing device, information processing is performed. If sensory data such as image signals and signals from measuring instruments are included in the processing target of the system, it is difficult to express these sensory data as rules based on symbols, and knowledge processing It is also difficult to improve the information processing ability and information processing accuracy by providing the device itself with a learning function.

Therefore, conventionally, as one method for solving the above problem, a configuration has been considered in which a neural network having a learning function is combined with a knowledge processing device represented by an expert system.

However, the conventional apparatus has a form in which the neural network is incorporated in the knowledge processing apparatus, and the knowledge processing apparatus and the neural network cannot be easily separated, and the program can be programmed according to the application. There was the inconvenience of having to recreate the whole thing.

[0005]

As described above, there has been no conventional efficient system in which the knowledge processing device and the neural network are integrated in a form that can be easily separated. Therefore, the entire program is created according to the application. There was the inconvenience of having to fix it.

Therefore, an object of the present invention is to provide an information processing system in which the knowledge processing device and the neural network are effectively integrated so that it is not necessary to change the entire program according to the application.

[0007]

In order to achieve the above object, the present invention provides a working memory capable of recording a symbol, referring to and deleting the recorded symbol, and a rule for rewriting the symbol recorded in the working memory. A production memory for recording, an inference engine for rewriting the recording of the working memory with reference to the recording of the working memory and the rules of the production memory, and a signal processing means for converting an input signal into a symbol and recording it in the working memory. An information processing system comprising signal output means for outputting a signal with reference to the record of the working memory, and control means for controlling the inference engine, the signal processing means, the signal output means, and the control means Characterized by configuring at least one by a method using a neural network To.

[0008]

At least one of the signal processing means for converting the input signal into a symbol and recording it in the working memory, the signal output means for outputting the signal with reference to the recording in the working memory, and the control means for controlling the inference engine. One is constructed by a method using a neural network. This makes it possible to construct a flexible and adaptive information processing system in which the knowledge processing device and the neural network are effectively integrated.

[0009]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of an information processing system according to the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing an embodiment of the information processing system of the present invention. The embodiment shown in FIG. 1 is constructed under the virtual control problem as shown in FIG. Therefore, first, the virtual control problem shown in FIG. 2 will be described.

In FIG. 2, there is a table T having casters driven by a motor between two persons h1 and h2, and a food F is placed on the table T.
The table T moves to three positions, a position X1 from the human h1, a position X2 from the human h2, and a position X0 intermediate between them. It is assumed that the human h1 can eat the food F on the table T when the table T comes to the position X1, and the human h2 can eat the food F on the table T when the table T comes to the position X2. Brain wave measuring devices S1 and S2 are attached to the humans h1 and h2, respectively.
The purpose of this control problem is to detect an annoyed person from the outputs of the electroencephalogram measuring devices S1 and S2, move the table T to the person and give food F on the table T to automatically annoy the annoyed person. The purpose is to build a controller for the table T to be solved.

Assume that the controller designer wants to incorporate the following four rules into the controller.

R1: When the human h1 is irritated, the table T is moved to the position X1.

R2: When the person h2 is irritated, the table T is moved to the position X2.

R3: When the table T is at the position X1, the table T is returned to the position X0.

R4: When the table T is at the position X2, the table T is returned to the position X0.

It is also assumed that the databases D1 and D2 showing the brain wave patterns of the annoyed state and the normal state for the humans h1 and h2 are known in advance.

FIG. 3 shows an example of a controller construction device for designing such a controller. This controller construction device comprises a neural network definition function 100, a production rule definition function 200, an inference control definition function 300, It has a conversion function 400. The controller designer designs the controller by calling each function.

The neural network defining function 100 is a function for constructing a desired neural network. Specifically, the neural network is constructed by specifying the name of the neural network and learning data. For example, the electroencephalogram measuring devices S1 and S2 shown in FIG.
When constructing the neural networks N1 and N2 that input the output of the above and outputs the symbols indicating whether the humans h1 and h2 are irritated or normal, respectively, while specifying the names of the neural networks N1 and N2, As the learning data, the database D1 showing the brain wave patterns of the human h1 in the annoyed state and the normal state and the database D2 showing the brain wave patterns of the human h2 in the annoyed state and the normal state are designated, and the neural network is specified. The definition function 100 is shown in FIG.
Generate a process as shown in.

Here, the process 1 for constructing the neural network N1
Reference numeral 01 refers to the database D1 to construct the neural network N1, and the neural network N2 configuration process 102 refers to the database D2 to construct the neural network N2. Also,
The neural network N1 construction process 101 and the neural network N2 construction process 102 call the neural network construction expert system 103 as needed when constructing the neural network N1 and the neural network N2.

The production rule definition function 200 is
It is a function of defining a production rule to be recorded in a production memory 133 described later. The production rule is a rule for rewriting the symbol recorded in the working memory 132 based on the symbol recorded in the working memory 132 described later. Working memory 13
The following three types of symbols are recorded in 2.

Simple symbol: A symbol that is referenced only by production rules.

Execution symbol: A symbol for executing a physical action, which is the production memory 132.
When recorded in, the physical action corresponding to this symbol is executed.

Interpretation symbol: a symbol that can be rewritten according to the result of monitoring the physical behavior of the target, and cannot be rewritten according to the production rule. In order to achieve the purpose of the control problem shown in FIG. 2, the production rule recorded in the production memory 133 is R1: h1 & X0 → + X1-X0 R2: h2 & X0 → + X2-X0 R3: X1 → + X0-X1 R4: X2 → It becomes + X0-X2. Here, + means recording in the working memory 132, and-means being deleted from the working memory 132.

Further, h1 and h2 are interpretation symbols, which are recorded in the working memory 132 when the state becomes irritated, and these symbols are deleted from the working memory 132 when the state becomes normal.

Further, X0, X1 and X2 are execution symbols, and when these symbols are recorded in the working memory 132, a control signal is sent to the table T so as to move to that position.

The inference control definition function 300 is a function for designating a conflict resolution strategy when two or more rules are simultaneously established in the production system 13 described later.
This conflict resolution strategy can be performed by designating a neural network, in which case the inference control defining function 300 can be replaced by the neural network defining function 100 as shown in FIG. Neural network definition function 100
Will be called again.

The conversion function 400 is a neural network definition function 1
, A neural network constructed by 00, a production rule defined by the production rule definition function 200,
This is a function to combine the conflict resolution strategies defined by the inference control definition function 300 into one. By this conversion function 400, the neural network constructed by the neural network definition function 100, the production rule defined by the production rule definition function 200, and the conflict resolution strategy defined by the inference control definition function 300 are put together into one figure. It is possible to construct a system that achieves the purpose of the control problem shown in FIG. 2, and FIG. 1 shows an embodiment of the information processing apparatus constructed in this way.

In FIG. 1, the interpreter 11 inputs the output of the electroencephalogram measuring device S1 for measuring the electroencephalogram of the human h1 and judges from the output of the electroencephalogram measuring device S1 whether the human h1 is irritated, If it is determined to be irritated, the interpretation symbol “h1” is recorded in the working memory 132 of the production system 13, and if not irritated, the interpretation symbol “h1” is deleted from the working memory 132 of the production system 13. Here, it is determined by the neural network N1 incorporated in the interpreter 11 whether or not the human h1 is irritated. This neural network N1 is constructed by the neural network defining function 100 as already described with reference to FIG.

Here, a specific configuration example of the neural network N1 will be described. It is assumed that the database D1 records brain wave sampling waveforms when the human h1 is in a normal state and when the person is irritated, and each waveform is sampled at n points. At this time, processing is performed by a three-layered neural network having n elements in the input layer and two elements in the output layer. One of the two elements of the output layer corresponds to a normal state and the other corresponds to a frustrated state. The sampling waveform of the electroencephalogram recorded in the database D1 is input to the neural network with n sampling points corresponding to n input elements in the input layer. If the input sampling waveform is normal, the output of the element corresponding to normal of the two output elements becomes high, and the output of the element corresponding to the other frustration becomes low. Train the neural network. If the input sampling waveform is an irritating one, the output of the element corresponding to the irritating one of the two output elements becomes high, and the output of the other element corresponding to the normal becomes low. To train the neural network. When an arbitrary sampling waveform recorded in the database D1 is input to the neural network, correct output can be obtained according to whether the waveform is normal or frustrated. Until
The above learning is repeated. The resulting neural network is
The sampling waveform obtained directly from the sensor S1 can be classified into two states, a normal state and an irritated state.

Further, the interpreter 12 inputs the output of the electroencephalogram measuring device S2 for measuring the electroencephalogram of the human h2, judges from the output of the electroencephalogram measuring device S2 whether the human h2 is irritated, and is irritated. If it is determined that the interpretation symbol “h” is stored in the working memory 132 of the production system 13,
2 ”is recorded, and if it is not irritating, the interpretation symbol“ h2 ”is deleted from the working memory 132 of the production system 13. Here, the determination whether the human h2 is irritating is incorporated in the interpreter 12. The neural network N2 is also constructed by the neural network defining function 100 as already described with reference to FIG.

The production system 13 has an inference engine 131, a working memory 132, and a production memory 133. In the production memory 133, the production rules R1: h1 & X0 → + X1-X0 R2 constructed by the above-described production rule definition function 200. : H2 & X0 → + X2-X0 R3: X1 → + X0-X1 R4: X2 → + X0-X2 is recorded.

Here, the production rule is IF
In the rule of the THEN ... format, the interpretation symbol “h1” indicating that the person h1 is irritated is recorded in the working memory 132 in the working memory 132, and the table T to be controlled is located at the central position X0. Execution symbol to position "X
If "0" is recorded, the execution symbol "X1" for moving the table T to the position h1 closer to the human h1 is recorded in the working memory 132, and the rule for deleting the execution symbol "X0" is shown. , The interpretation symbol “h2” indicating that the human h2 is irritated in the working memory 132
Is recorded and an execution symbol “X0” for recording the table T to be controlled at the central position X0 is recorded, the table T is stored in the working memory 132 as a human h2.
A rule for recording the execution symbol "X2" to be moved to the closer position X2 and deleting the execution symbol "X0" is shown as R3.
If the execution symbol "X2" for locating the table T to be controlled at the position X2 near the human h2 is recorded in the working memory 132, the working memory 13
2 shows a rule for recording the execution symbol "X0" for moving the table T to the central position X0 and deleting the execution symbol "X2", and R4 is a table to be controlled in the working memory 132. When the execution symbol "X1" located at the position X1 near the h1 is recorded, the execution symbol "X0" for moving the table T to the central position X0 is recorded in the working memory 132, and the execution symbol "X1" is deleted. The rule to do is shown.

The inference engine 131 constantly monitors the symbols recorded in the working memory 132, and executes the rule if any of the rules described in the production memory 133 satisfies the condition part.

The rule monitor 15 monitors the working memory 132 and the production memory 133 of the production system 13, and transmits the number of the executed rule and the state of the working memory 132 to the inference control neural network 16.

The inference control neural network 16 has a plurality of elements corresponding to the respective rules recorded in the production memory 133, and the strength of the connection between the respective elements corresponding to the contents transmitted from the rule monitor 15. To change. Then, when the items recorded in the working memory 132 simultaneously satisfy the condition parts of two or more rules among the rules described in the production memory 133, one of the rules satisfying the condition parts is satisfied. A rule is selected and the result of this selection is transmitted to the inference engine 141 to perform a conflict resolution strategy decision.

When the condition parts of two or more rules among the rules described in the production memory 133 are satisfied, the inference control neural network 16 outputs the two or more rules whose condition parts are satisfied. The process of selecting one rule will be further described. The four rules described in the production memory 133 are R1, R2, R3, and R4, and the inference control neural network 16 has four elements corresponding to these four rules. The rule monitor 15 outputs the output of the element corresponding to the rule while there is no need for conflict resolution, that is, while one of the rules described in the production memory 133 satisfies the condition part. Make it higher Then, when the condition parts of two or more rules are satisfied, the inference control neural network 16 is operated. The inference control neural network 16 exchanges signals between the elements and eventually becomes a balanced state, and the output of each element becomes constant. At this time, the outputs of the elements corresponding to the two or more rules satisfying the condition part are compared to find the element having the largest output. The conflict is resolved by selecting the rule corresponding to the element with the maximum output.

In the inference engine 131, the items recorded in the working memory 132 are stored in the production memory 133.
If the condition part of one of the rules described in 1 is satisfied, the rule is executed and the condition parts of two or more of the rules described in the production memory 133 are satisfied. Then, the inference control neural network 16 selects one rule by the method described above and executes the selected rule.

The control controller 14 monitors the working memory 132 of the production system 13, forms a control signal corresponding to the execution symbol when the execution symbol is recorded in the working memory 132, and outputs this control signal to the table T. It is sent to a motor (not shown) to be driven.

In FIG. 1, solid line arrows indicate the flow of control signals, and dotted line arrows indicate the flow of data.

With such a configuration, control that achieves the purpose of the control problem shown in FIG. 2 can be realized.

In the configuration shown in FIG. 1, regarding the inference of the inference engine 131 of the production system 13, when there is a high probability that there is no rule that satisfies two or more condition parts at the same time, the rule monitor 15 and the inference are performed. The configuration of the control neural network 18 can be omitted.

FIG. 6 shows another embodiment of the present invention constructed as described above.

Further, in FIG. 1, a neural network is incorporated in the control controller 14, and the neural network is configured to convert the execution symbols recorded in the working memory 132 into the control signals of the table T. Good.

FIG. 7 shows still another embodiment of the present invention constructed as described above, and the control controller 14
The neural network N3 is incorporated in the neural network, and the control controller 14 monitors the working memory 132 of the production system 13, and when the execution symbol is recorded in the working memory 132, the control signal corresponding to the symbol is output to the neural network. It is formed by using N3, and this control signal is transmitted to the table T
To a motor (not shown) that drives the.

That is, the execution symbols of the rules described in the production memory 133 shown in FIG. 7 are X0 and X.
There are only three, 1 and X2, but when adding execution symbols such as slowly returning the table T to the center as X0s and quickly returning the table T to the center as X0f,
The control signal sent to the motor driving the table T needs to be changed depending on whether the execution symbol is X0s or X0f. In this case, the neural network N3 converts the execution symbols into control signals for the motors driving the table T.

With such a configuration, it is possible to solve a wider range of problems flexibly and adaptively.

[0048]

As described above, according to the present invention,
Since the information processing system is constructed by effectively integrating the neural network in the intelligent processing device, it is possible to construct a flexible and adaptive information processing system.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of an information processing system of the present invention.

FIG. 2 is a diagram illustrating a virtual control problem that is a target of the embodiment of FIG.

FIG. 3 is a block diagram showing an example of a controller construction device that designs a controller for solving the virtual control problem shown in FIG.

FIG. 4 is a block diagram showing an example of a neural network definition process generated by the neural network definition function shown in FIG.

5 is a block diagram showing an example of an inference control definition process generated by the inference control definition function shown in FIG.

FIG. 6 is a block diagram showing another embodiment of the information processing system of the present invention.

FIG. 7 is a block diagram showing still another embodiment of the information processing system according to the present invention.

[Explanation of symbols]

 11, 12 Interpreter 13 Production system 14 Control executor 15 Rule monitor 16 Inference control neural network 100 Neural network definition function 200 Production rule definition function 300 Inference control definition function 400 Conversion function N1, N2, N3 Neural network S1 , S2 EEG analyzer T table F Food h1, h2 Human

Claims (1)

[Claims] 1. A working memory capable of recording a symbol and referring to and deleting the recorded symbol, a production memory for recording a rule for rewriting the symbol recorded in the working memory, a recording in the working memory and the production memory An inference engine that rewrites the record of the working memory by referring to the rule of 1., a signal processing unit that converts an input signal into a symbol and records the symbol in the working memory, and a signal that outputs a signal by referring to the record of the working memory. In an information processing system including output means and control means for controlling the inference engine, at least one of the signal processing means, the signal output means, and the control means is configured by a method using a neural network. Information processing system characterized by.