GB2259155A - Food trolley controller - Google Patents

Abstract

The state of an object to be controlled eg a food trolley T is monitored and stored as a symbol in a working memory 132 of an information processing means 13. An inference engine 13 rewrites the stored symbols according to a set of rules R1 to R4 stored in a production memory 133 and the trolley is controlled according to the contents of the working memory. The control system includes a neural network 16 and can be used to cause the trolley to move to a person h1, h2 who is detected as being irratable. <IMAGE>

Description

"SUBJECT CONTROLLING SYSTEM" This invention relates to a subject controlling system for controlling the subject such as a position of an objective member or a temperature, the system being constructed by combining a neural network with a knowledge processing device such as an expert system using a production system.

Generally, in a knowledge processing device represented by an expert system, knowledge is expressed by use of symbols in principle, and therefore, when a system for processing information and controlling a subject by use of the result of the information processing is constructed by use of such a knowledge processing device and if sensor data such as a signal from a measuring device or an image signal is included in an objective information to be processed in the information processing, it is difficult to express the sensor data as a rule based on the symbols. It is also difficult to design the knowledge processing device to have the learning function in order to enhance the information processing ability, information processing precision and the like.Further, it is desired to develop a subject controlling system capable of easily and rapidly responding to modification of applications.

An object of this invention is to provide a subject controlling system which is constructed by effectively combining a neural network with a knowledge processing device so as to effect the information processing without necessity of changing the entire program according to applications.

The above object can be attained by a subject controlling system which comprises: means for outputting an input signal indicating the state of a to-be-controlled subject; information processing means including a working memory in which symbols can be stored and stored symbols can be referred to and deleted, a production memory for storing rules according to which symbols stored in the working memory are rewritten, an inference engine for rewriting the stored content of the working memory by referring to the rule of the production memory and the stored content of the working memory, signal processing means for converting the input signal into a symbol and storing the converted symbol into the working memory, signal outputting means for outputting a signal by referring to the stored content of the working memory, control means for controlling the inference engine, and a neural network contained in at least one of the signal processing means, signal outputting means and control means; and means for controlling the to-be-controlled subject according to an output signal from the signal outputting means.

This invention can be more fully understood from the following detailed description when taken in conjunction with the accompanying drawings, in which: Fig. 1 is a block diagram showing one embodiment of a subject controlling system according to this invention; Fig. 2 is a view for explaining a to-be-controlled subject in the embodiment of Fig. 1; Fig. 3 is a block diagram showing an example of a controller constructing device for designing a controller for solving a control problem shown in Fig. 2; Fig. 4 is a block diagram showing an example of a neural network defining process created by the neural network defining function; Fig. 5 is a block diagram showing an example of an inference control defining process created by the inference control defining function;; Fig. 6 is a block diagram showing another embodiment of the subject controlling system according to this invention; Fig. 7 is a block diagram showing still another embodiment of the subject controlling system according to this invention; Fig. 8 is a diagram showing an example of a neural network used in this invention; and Fig. 9 is a diagram showing another example of an inference controlling neural network used in this invention.

There will now be described an embodiment of a subject controlling system according to this invention with reference to the accompanying drawings.

Fig. 1 is a block diagram showing one embodiment of the system of this invention. The embodiment shown in Fig. 1 is constructed based on the control problem as shown in Fig. 2. Therefore, the control problem shown in Fig. 2 is first explained.

As shown in Fig. 2, there is a table T with casters which is driven by a motor between two persons hl and h2 and food F is placed on the table T. The table T is moved between three positions, that is, a position X1 set near the person hl, a position X2 set near the person h2 and an intermediate position X0. The person hl can take and eat the food F when the table T comes to the position X1 and the person h2 can take and eat the food F when the table T comes to the position X2.

Electroencephalographs S1 and S2 are respectively attached to the persons hl and h2. An object of the control problem is to construct a controller for controlling the table T by detecting one of the two persons who is irritated based on outputs of the electroencephalographs S1 and S2 and moving the table T towards the person who is irritated so that the person can eat the food on the table and can be automatically relieved from irritation.

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

R1: Move the table T to the position Xl if the person hl becomes irritated.

R2: Move the table T to the position X2 if the person h2 becomes irritated.

R3: Return the table T to the position XO if the table T lies on the position X1.

R4: return the table T to the position XO if the table T lies on the position X2.

Further, assume that data bases Dl and D2 (to be described later with reference to Fig. 4) indicating the brain wave patterns of the respective persons hl and h2 in the irritated state and normal state are previously known.

Fig. 3 shows an example of a controller constructing device for designing the above controller, and the controller constructing device includes a neural network defining function 100, production rule defining function 200, inference control defining function 300 and converting function 400. The designer of the controller designs the controller by adequately reading out the respective functions.

The neural network defining function 100 is a function of constructing a desired neural network.

More specifically, it is a function of constructing a neural network by specifying the name and learning data of the neural network. For example, when neural networks N1 and N2 which output symbols indicating whether the persons hl and h2 are in the irritated state or normal state in response to outputs of the electroencephalographs S1 and S2 are constructed, the neural network defining function 100 specifies the names of the neural networks N1 and N2 and specifies data base D1 indicating the brain patterns of the person hl in the irritated and normal states and data base D2 indicating the brain patterns of the person h2 in the irritated and normal states as learning data to create processes as shown in Fig. 4.

A neural network Nl constructing process 101 is created to construct a neural network N1 with reference to the data base Di and a neural network N2 constructing process 102 is created to construct a neural network N2 with reference to the data base D2. Further, the neural network N1 constructing process 101 and neural network N2 constructing process 102 read out a neural network constructing expert system 103 as required when constructing the neural network N1 and neural network N2.

The production rule defining function 200 is a function of defining the production rule to be stored in a working memory 133 which will be described later. The production rule is a rule for rewriting symbols to be stored in a working memory 132 which will be described later based on the symbols stored in the working memory 132. The following three types of symbols are stored in the working memory 132.

Simple Symbol: a symbol which is referred to only by the production rule.

Execution Symbol: a symbol used for executing a physical action. When the symbol is stored in the production memory 132, a physical action corresponding to the symbol is executed.

Interpretive Symbol: a symbol which can be rewritten according to the monitoring result of the physical action of the object and which cannot be rewritten by the production rule.

In order to attain the object of the control problem shown in Fig. 2, production rules stored in the production memory 132 by symbols are set as follows: R1: hl & 0 < +X1-XO R2: h2 & 0 o +X2-X0 R3: X1 e +X0-X1 R4: X2 < +X0-X2 where "+" indicates that a symbol is written into the working memory 132 and "-" indicates that a symbol is deleted from the working memory 132.

Further, hl and h2 are interpretive symbols which are stored in the working memory 132 when the irritated state is detected and deleted from the working memory 132 when the normal state is detected.

X0, X1 and X2 are execution symbols, and when the execution symbol is stored into the working memory 132, a control signal is supplied to the table T to move to a corresponding position.

The inference control defining function 300 is a function of specifying a competition canceling strategy when two or more rules are simultaneously set up in a production system 13 which will be described later. The competition canceling strategy can be effected by specifying the neural network, and in this case, the inference control defining function 300 can be replaced by the neural network defining function 100, and at this time, the neural network defining function 100 is read out again.

The converting function 400 is a function of putting together the neural network constructed by the neural network defining function 100, the production rule defined by the production rule defining function 200 and the competition canceling strategy defined by the inference control defining function 300. A system for attaining the object of the control problem shown in Fig. 2 can be constructed by putting together the neural network constructed by the neural network defining function 100, the production rule defined by the production rule defining function 200 and the competition canceling strategy defined by the inference control defining function 300 by use of the converting function 400. Fig. 1 shows an example of an information processing device thus constructed.

In Fig. 1, an interpreting unit 11 receives an output of the electroencephalograph S1 for measuring the brain wave of the person hl via an electrode attached to the front end of a cable cl to periodically determine whether the person hl is irritated or not based on the brain wave output from the electroencephalograph S1.

When it is determined that the person is irritated, the interpreting unit 11 outputs an interpretive symbol "hl" so as to store the interpretive symbol into the working memory 132 of the production system 13, and when it is determined that the person is not irritated, the interpretive symbol "hl" is deleted from the working memory 132 of the production system 13. In this case, determination whether the person hl is irritated or not is made by the neural network N1 incorporated into the interpreting unit 11. The neural network N1 is constructed by the neural network defining function 100 as is already explained with reference to Fig. 4.

Now, an example of the concrete construction of the neural network N1 is explained. Sampling waveforms of brain waves generated when the person hl is in the normal state and irritated state are contained in the data base D1 and each waveform is sampled at four points, for example. In this case, the process is effected by use of the three-level hierarchical neural network N1 having four elements I1 to I4 in the input level, three elements M1 to M3 in the intermediate level and two elements 01 and 02 in the output level as shown in Fig. 8.The element 01 which is one of the two elements in the output level is set to correspond to the normal state and the other element 02 is set to correspond to the irritated state. The sampling waveform of the brain wave contained in the data base D1 is input to the neural network with the four sampling points of the sampling waveform set to correspond to the four input elements I1 to I4 in the input level. The neural network N1 is caused to learn to set an output of the element 01 which is one of the two output elements and corresponds to the normal state to a high level and set an output of the other output element 02 corresponding to the irritated state to a low level when the input sampling waveform corresponds to the normal state. Further, the neural network N1 is caused to learn to set an output of the element 02 which is one of the two output elements and corresponds to the irritated state to a high level and set an output of the other output element 02 corresponding to the normal state to a low level when the input sampling waveform corresponds to the irritated state. When a desired sampling waveform contained in the data base D1 is input to the neural network N1, the above learning operation is repeatedly effected until a correct output can be derived according to whether the input waveform corresponds to the irritated state or normal state. The neural network N1 thus constructed can classify sampling waveforms directly supplied from the sensor S1 into two groups of normal state and irritated state.The interpreting unit 11 outputs an interpretive symbol "hl" which is stored in the working memory 132 or deleted from the working memory 132 according to the output of the neural network N1.

Likewise, an interpreting unit 12 receives an output of the electroencephalograph S2 for measuring the brain wave of the person h2 to determine whether the person h2 is irritated or not based on the brain wave output from the electroencephalograph S2. When it is determined that the person is irritated, the interpreting unit 12 outputs an interpretive symbol "h2" so as to store the interpretive symbol into the working memory 132 of the production system 13, and when it is determined that the person is not irritated, the interpretive symbol "h2" is deleted from the working memory 132 of the production system 13. In this case, determination whether the person hl is irritated or not is made by the neural network N2 incorporated into the interpreting unit 12.The neural network N2 is constructed by the neural network defining function 100 as is already explained with reference to Fig. 4 and actually constructed as shown in Fig. 8.

The production system 13 includes an inference engine 131, working memory 132, and production memory 133 and the following production rules R1 to R4 constructed by the production rule defining rule 200 are stored in the production memory 133.

R1: hl & O 3 +X1-XO R2: h2 & O e +X2-XO R3: X1 > +XO-X1 R4: X2 3 +XO-X2 In this case, the above production rule is a rule formed in the format of "IF---THEN---", R1 indicates a rule for storing an execution symbol "X1" for moving the table T to the position X1 near the person hi into the working memory 132 and deleting an execution symbol "XO" for moving the table T which is an object to be controlled to the central position XO if the interpretive symbol "hl" indicating that the person hl is irritated is stored in the working memory 132 and an execution symbol "XO" is stored, R2 indicates a rule for storing an execution symbol "X2" for moving the table T to the position X2 near the person h2 into the working memory 132 and deleting the execution symbol "XO" if the interpretive symbol "h2" indicating that the person h2 is irritated is stored in the working memory 132 and the execution symbol "XO" for moving the table T which is an object to be controlled to the central position XO is stored, R3 indicates a rule for storing the execution symbol "XO" for moving the table T to the central position XO into the working memory 132 and deleting the execution symbol "X2" if the execution symbol "X2" for moving the table T which is an object to be controlled to the position X2 near the person h2 is stored in the working memory 132, and R4 indicates a rule for storing the execution symbol "XO" for moving the table T to the central position XO into the working memory 132 and deleting the execution symbol "X1" if the execution symbol "X1" for moving the table T which is an object to be controlled to the position Xl near the person hl is stored in the working memory 132.

The inference engine 131 always monitors the symbols stored in the working memory 132, and if a rule whose conditional part is satisfied is detected in the rules described in the production memory 133, then the rule is executed.

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

The inference control neural network 16 has a plurality of elements corresponding to the respective rules stored in the production memory 133 and changes the strength of connection between the elements according to the content transmitted. When the item stored in the working memory 132 simultaneously satisfies the conditional parts of two or more of the rules described in the production memory 133, the competition canceling strategy determination for selecting one of the rules whose conditional parts are satisfied and transmitting the result of selection to the inference engine 141 is effected.

The process in which the inference control neural network 16 selects one of the two or more rules whose conditional parts are satisfied when the conditional parts of two or more of the rules described in the production memory 133 are simultaneously satisfied is further explained. The four rules R1, R2, R3 and R4 are described in the production memory 133 and four elements rl to r4 respectively corresponding to the four rules are provided in the inference control neural network 16.

When the competition canceling operation is not necessary, that is, when the conditional part of only one of the rules described in the production memory 133 is satisfied, the rule monitor 15 sets an output of the element corresponding to the rule to a high level. When the conditional parts of two or more rules are satisfied, the inference control neural network 16 is operated. The inference control neural network 16 transfers a signal between the elements rl to r4, then an equilibrium state can be set up, and the outputs of the elements rl to r4 are made stable. At this time, the outputs of the elements rl to r4 corresponding to the two or more rules whose conditional parts are satisfied are compared with each other to detect the element having the largest output.Thus, the competition canceling operation is effected by selecting the rule corresponding to the element having the largest output.

Fig. 9 shows another example of the inference control neural network shown in Fig. 1. The neural network 60A has a hierarchical structure having an input level constructed by four input neural elements I1 to I4 corresponding to the four rules R1 to R4 stored in the production memory 133 in the embodiment of Fig. 1, an output level constructed by four output neural elements 01 to 04, and intermediate level constructed by three intermediate neural elements Ml to M3.

When inputs are supplied to the input elements I1 to I4 in a condition that an input which satisfies the conditional part of at least one of the rules R1 to R4 is set to Ii=l (i = 1 to 4) and an input which does not satisfy the conditional part of any one of the rules is set to Ii=O (i = l to 4), outputs corresponding to the inputs are derived from the four output elements 01 to 04, and if the largest output is derived from an output element Oj (j = 1 to 4), a rule Rj corresponding to the output element Oj is selected.

When the item stored in the working memory 132 satisfies the conditional part of one of the rules described in the production memory 133, the inference engine 131 executes the rule, and when the item satisfies the conditional parts of two or more rules described in the production memory 133, one of the rules is selected according to the above method by means of the inference control neural network 16 and then the inference engine 131 executes the selected rule.

A control executing unit 14 monitors the working memory 132 of the production system 13 and when an execution symbol is stored into the working memory 132, it creates a control signal corresponding to the symbol, supplies the control signal to a motor controller MC so as to control the forward rotation, reverse rotation and interruption of a motor M for moving the table T.

In Fig. 1, arrows indicated by solid lines indicate the flow of the control signal and arrows indicated by broken lines indicate the flow of data.

With the above construction, the control operation for attaining the object of the control problem shown in Fig. 2 can be realized.

In the construction shown in Fig. 1, if the possibility that the conditional parts of two or more rules can be simultaneously satisfied is low in view of the inference of the inference engine 131 of the production system 13, the construction of the rule monitor 15 and inference control neural network 16 may be omitted.

Fig. 6 shows another embodiment of this invention thus constructed.

It is also possible to incorporate a neural network into the control executing unit 14 shown in Fig. 1 and convert an execution symbol stored in the working memory 132 to a control signal for controlling the table T by use of the incorporated neural network.

Fig. 7 shows still another embodiment of this invention thus constructed and a neural network N3 is incorporated into a control executing unit 14a. The control executing unit 14a monitors the working memory 132 of the production system 13, and when an execution symbol is stored into the working memory 132, it creates a control signal corresponding to the symbol by use of the neural network N3 and supplies the control signal to a motor M for driving the table T.

That is, execution symbols of the rules described in the production memory 133 shown in Fig. 7 are only three execution symbols XO, Xl and X2, but if an execution symbol XOs for slowly returning the table T to the central position and an execution symbol XOf for rapidly returning the table T to the central position are additionally provided, it becomes necessary to change the control signal supplied to the motor M for driving the table T according to whether the execution symbol is XOs or XOf. In this case, the execution symbol is converted to the control signal for the motor M for driving the table T by means of the neural network N3.

With the above construction, it becomes possible to solve problems in a wider range in a flexible and adaptive manner.

Though the embodiments shown in Figs. 1, 6 and 7 include the table T as a subject to be controlled according to the present invention, a physical quantity such as a room temperature can be the subject to be controlled.

As described above, according to this invention, since a subject controlling system is constructed by effectively combining a neural network with a knowledge processing device so as to effect the information processing, it can effect the information processing in a flexible and adaptive manner.

Claims (7)

Claims:

1. A subject controlling system comprising: means for outputting an input signal indicating the state of a to-be-controlled subject; information processing means including a working memory in which symbols can be stored and stored symbols can be referred to and deleted, a production memory for storing rules according to which symbols stored in said working memory are rewritten, an inference engine for rewriting the stored content of said working memory by referring to the rule in said production memory and the stored content of said working memory, signal processing means four converting the input signal into a symbol and storing the converted symbol into said working memory, signal outputting means for outputting a signal by referring to the stored content of said working memory, engine control means for controlling said inference engine, and a neural network contained in at least one of said signal processing means, signal outputting means and control means; and means for controlling the to-be-controlled subject according to an output signal from said signal outputting means.

2. A system according to claim 1, wherein said engine control means includes competition canceling means having means for monitoring whether or not the content of said working memory simultaneously satisfies the conditional parts of a plurality of rules in said production memory and a neural network for effecting the inference control to select one of said plurality of rules and output the selected rule to said inference engine.

3. A system according to claim 2, wherein said inference controlling neural network includes neural elements of a number corresponding to the number of rules stored in said production memory, and means for connecting said neural elements to one another so as to change the strength of connection between said neural elements according to an output of said monitoring means.

4. A system according to claim 1, wherein said neural network has a hierarchical structure having an input level including input neural elements of a number corresponding to the sampling number of an input signal, an output level including output elements of a number corresponding to the number of necessary answers, and an intermediate level including a preset number of intermediate neural elements connected between said input and output levels.

5. A subject controlling system comprising: means for outputting an input signal indicating the state of a to-be-controlled subject; information processing means including a working memory in which symbols can be stored and stored symbols can be referred to and deleted, a production memory for storing rules according to which symbols stored in said working memory are rewritten, an inference engine for rewriting the stored content of said working memory by referring to the rule in said production memory and the stored content of said working memory, signal processing means for converting the input signal into a symbol and storing the converted symbol into said working memory, signal outputting means for outputting a signal by referring to the stored content of said working memory, and a neural network contained in at least one of said signal processing means and signal outputting means; and means for controlling the to-be-controlled subject according to an output signal from said signal outputting means.

6. A system according to claim 5, wherein said neural network has a hierarchical structure having an input level including input neural elements of a number corresponding to the sampling number of an input signal, an output level including output elements of a number corresponding to the number of necessary answers, and an intermediate level including a preset number of intermediate neural elements connected between said input and output levels.

7. A subject controlling system, substantially as hereinbefore described with reference to the accompanying drawings.