JPH06348675A - Neuro-computer application equipment and machinery provided with the neuro-computer application equipment - Google Patents

Abstract

PURPOSE:To provide a practical neuro-computer capable of performing complicated processings such as the optimum operation of a plant or the like which was actually difficult to be made practicable before. CONSTITUTION:This equipment is composed of a generation means 11 for generating neurons and synapses, a means for connecting the generated neurons and synapses and constructing a neural network and a driving means 13 for inputting a feature amount and teacher signals to the neural network 14 and performing learning and is provided with the means 15 for exchanging the optional neurons constituting the neural network with the other neurons with different input/output characteristics or optionally connecting and disconnecting the neurons.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a neuro computer having a learning function capable of performing optimum operation of a plant having a complicated equipment configuration, and particularly, it is simulated by a simulation program operating on a Neumann type computer. The structure of the created neuro computer.

[0002]

2. Description of the Related Art A neurocomputer is similar to the nervous system of an organism in that it mimics the mechanism of the nervous system of an organism by engineering, and signals are transmitted in a massively parallel manner among many neurons corresponding to nerve cells. It is a computer that is trying to realize various information processing. There are various possibilities that cannot be realized by information processing by a Neumann computer that performs serial processing, and they have been recently developed.

It is said that it is difficult for the neurocomputer at the present stage to make the neurons themselves, which are the units of nerve cells, into hardware. Therefore, a neural computer is generally simulated by implementing a neural network model by a simulator, that is, a computer program, and processing the program using a conventional Neumann type computer itself.
Many conventional manuals are available on the market for neurocomputers, for example, see Shunichi Amari's "Neurocomputer" published by Yomiuri Shimbun.

The structure of a neuro computer simulated on a Neumann type computer by a conventional simulation program and its learning method will be described below.

FIG. 9 shows an S layer 4 composed of sensor neurons.
This is an example of a hierarchical neural network configuration of three R layers each consisting of one response neuron. Here, FIG.
, The sensor neurons of the S layer are s1 to s4, R
The response neurons of the layer are r1 to r3, and the input is x
1 to x4 and outputs u1 to u3. The thresholds of the response neurons are θ1 to θ3.

In the conventional neurocomputer, the relationship between the input and the output is all expressed by a matrix. That is,
For example, the input / output load relationship is expressed by a matrix relational expression as shown in the table below.

[Table 1] To obtain the output vector from the input vector, input X,
The weight matrix W, the threshold value θ, and the output U are respectively represented by a vector (matrix) as follows.

[Equation 1] The output vector U is expressed by the following equation. U = f (W * X * (theta)) Here, f is a function showing the input-output characteristic of a neural network.

In the case of applying input data and a teacher signal to such a neural network for learning, a conventional neurocomputer uses a method of recognizing a pattern vector. In this method, the data to be recognized is directly input as a pattern, and the algorithm of the neuro computer with high recognition ability is used to recognize all in the learning process of the neuro computer. Specifically, the initial value of the synapse coupling weight matrix W is given by, for example, a random number. Then, the input vector is given and the output vector is calculated. The difference ΔW of the weight matrix W is calculated from the learning algorithm by comparing the output result with the teacher signal. Then, the weight matrix is rewritten to W-ΔW for correction. This process is performed for all the sets of input signals and teacher signals, and this combination is repeated several times or several tens of times, and the synapse coupling weight matrix W becomes the teacher signals for all inputs. It is modified in the matching direction (learn process).

[0008]

However, in such a conventional learning algorithm for a neurocomputer, a computer having an ultra-high speed and a large capacity is used except when the input pattern clearly shows the characteristics and a large amount of learning data exists. It was unlikely that learning would succeed even when using. That is,
In order to obtain a neurocomputer capable of accurately operating a plant having a complicated equipment configuration, the weight matrix W does not converge no matter how many times the learning process of correcting the weight matrix W of the neural network is repeated.
In many cases, the optimum load matrix W cannot be obtained. If the learning process does not converge, or if the unlearned data cannot be correctly recognized, the normalization of the input data is changed or the composition of the network is changed.
Changes are trial-and-error, as networks generally get quite large and treat the interior as a black box.
This is extremely unlikely to succeed.

Even with the conventional learning algorithm, automatic recognition,
The learning process may be successful when performing category classification or the like. However, in the conventional learning algorithm, in many cases, it is practically unusable for a slightly complicated problem such as finding a neuro computer that can optimally operate a plant having a complicated device configuration.

In view of the above-mentioned problems of the prior art, the present invention provides a neurocomputer which can be practically used for solving complicated problems such as optimum operation of a plant having a complicated equipment configuration within the range of the conventional learning algorithm. The purpose is.

[0011]

According to the present invention, a generating means for generating neurons and synapses, a means for connecting the generated neurons and synapses to construct a neural network, a feature quantity and a teacher for the neural network. Drive means for inputting and learning a signal, and exchanging any neuron constituting the neural network with another neuron having different input / output characteristics,
Alternatively, it is characterized in that a means for arbitrarily connecting and disconnecting the neurons is provided.

Further, in the present invention, the synapse includes at least information on a connection relation with input / output neurons and load data, and the neurons are composed of sensor neurons and response neurons, and at least each of the synapses is connected to the synapse. It is characterized in that the information and data of the connection relation and the response neuron are provided with a program of a learning function, and the data of the synapse weight is rewritten by the learning function of the response neuron.

[0013]

The present invention is provided with a generation means for generating neurons and synapses having arbitrary characteristics and a means for constructing a neural network, and exchanges neurons in the neural network with other neurons having different input / output characteristics, or disconnects between neurons. Since the means for enabling connection is provided, the input / output characteristics of each neuron are changed from the internal state of the neurocomputer and the causal relationship with the output,
The network structure can be changed locally.
Therefore, it becomes possible for a neurocomputer developer to revise a defect part of the neural network to a neuron having an optimal characteristic in the middle of the learning process, and the learning process can be converged in a short period of time. It is possible to realize a practical neurocomputer that requires a solution for a complicated system.

[0014]

1 is a block diagram of a neurocomputer according to an embodiment of the present invention. This neurocomputer has a generating means 11 for generating neurons and synapses.
A means 12 for constructing a neural network 14 by connecting the generated neurons and synapses, and a drive means 13 for inputting a feature amount and a teacher signal to the neural network for learning. Each neuron generated by the generating means 11 has a connection relation with input / output synapses, data such as threshold values, input / output characteristics, a learning program, etc., and each synapse generated by the generating means 11 is input / output. It has data such as the connection relations with the neurons of and the weights. Further, there is provided means 15 for exchanging an arbitrary neuron constituting the neural network with another neuron having different input / output characteristics, or arbitrarily connecting and disconnecting the neurons.

In the present invention, object-oriented programming is introduced into the software that constitutes the neurocomputer. In object-oriented programming, an encapsulation of a data structure and the behavior (procedure) of that data is defined as an object, which is the minimum unit of software. Object templates are called classes, and related objects can have inheritance relationships. The software is driven by exchanging messages between objects. The object-oriented modeling is performed by faithfully mapping an actual "object" to an object on a computer. With the introduction of object-oriented programming, the neurons and synapses that encapsulate the data structure (attribute) and the behavior (method) of that data are defined as objects, and these are the minimum units that make up the neural network.

In the neurocomputer of the present invention, a class (mold) for generating each object is defined as shown in FIG. Here, the synapse connection weight (weight) which is the essence of the neurocomputer is held as synapse data (attribute). A synapse class (mold) is a pointer (input neuron, output neu) that indicates the connection relationship between the input and output neurons to which the synapse connects.
ron) as data (attribute). Further, it has an input / output function characteristic (g) as a behavior (method). This input / output characteristic (g) normally uses a linear function. On the other hand, the class (mold) of a neuron is data (attribute) as a threshold value (θ), a pointer (input synapse, output synapse) indicating the connection relationship between the input and output synapses connected to the neuron, and the output value ( u) and has a function characteristic (f) of input / output and a learning function (learn) program as a method. Therefore, the synapse itself holds the connection weight (weight), and the input / output characteristic (f), threshold value (θ), and learning function program (lear).
n) is held independently by the neuron itself.

FIG. 3 is an explanatory diagram showing the construction of a neural network according to an embodiment of the present invention. FIG. 3A shows the inheritance structure of classes (molds). In the neural network 14 shown in FIG. 1, sensor neurons of the S layer and R
In the response neurons of the layers, the behavior of the neurons is different. For example, the learning neurons are not required for the sensor neurons of the S layer, but the learning algorithms are required for the response neurons of the R layer. On the other hand, input / output characteristics are required for both sensor neurons and response neurons.
Specifies input and output functions. Therefore, the neuron
It is roughly divided into sensor neurons and response neurons,
There is a learning algorithm (learn) and a difference in input / output characteristics. These are generated as an inheritance structure from a common class (mold) with a difference in data and method.

Generally, in the class inheritance structure in object-oriented programming, common attributes and methods between objects are defined in a superordinate class, and only different parts are defined in a subordinate class. Since the lower class inherits the attributes and methods of the upper class, the upper class represents an abstract concept, and the lower class becomes more specific. The neurocomputer of the present invention is characterized in that synapses and neurons are defined in the highest class, and the neurons are defined in the following higher classes as sensor neurons and response neurons. Synapses have no subclass because they use common ones in the network. Therefore,
The neuron has an inheritance structure as shown in FIG.

In the neurocomputer of the present invention, a neural network 14 is constructed by means 12 for generating synapse and neuron objects from each class (mold) and linking the generated neurons and synapses. Further, the neural network 14 is provided with drive means 13 for learning by inputting a feature amount and a teacher signal, and the neural network is driven by message transmission between objects or by calling a method, and output vector calculation, teacher signal Learning processes such as comparison, calculation of the synapse coupling load ΔW, correction of the synapse coupling load W, etc. are performed (see FIG. 1).

The formation of the neural network is shown in FIG.
As shown in (B) and (C), the two-step network construction means 12 is used. 1) First, as shown in FIG. 3B, neurons and synapses having predetermined data and methods, which are objects, are generated from each class (mold). 2) Next, as shown in FIG. 3 (C), the synapses are linked according to a network structure assumed in advance. A random number or a constant value is input to the synapse coupling weight W and the like in the initial state.

The network is driven according to the procedure shown in FIG. 4 by the driver means 13 for learning by inputting the feature amount and the teacher signal. Step 1) Input vector x1 to the sensor neuron s1
Enter. Step 2) s1 outputs a value according to the input / output characteristic f, which is a method, to the synapse. Step 3) The synapse uses the input value and the connection weight W as an attribute to calculate the value according to the input / output characteristic (g) as r1.
Output to. Step 4) Similarly, all synapses linked to r1 output to r1. Step 5) r1 sums the input values and outputs u1 according to the input / output characteristic (f).

In learning, the output (u) and the teacher signal are arithmetically compared based on the learning program (learn) of the method of each neuron. The connection weight (ΔW) of each synapse, which is the difference between the teacher signal and the output value (u), is calculated,
The data of the difference (ΔW) of the connection weight is passed from the neuron to each synapse, and the new connection weight (W-
The data is rewritten to ΔW). Since the method can be defined for each class, if a class with different input / output characteristics is prepared in advance and a neuron object is generated from that class, learning output based on different operations for each neuron is possible. become.

Generation of neurons and synapses, and construction of a network by their links are all provided with means 15 that can be changed at any time (see FIG. 1).
Therefore, even after the neurocomputer has performed learning by comparing a set of inputs with a teacher signal, a neuron is exchanged with another class of neuron having different input / output characteristics, or a neuron is added to the network. It is possible to dynamically change the configuration of.

For example, the sensor neuron forming the input layer (S layer) of the network has a role of transmitting the input value as it is to the neuron of the response layer (R layer). Therefore, normally, a sensor neuron having a linear characteristic (f) is used. Here, as a result of observing and analyzing the internal state of the neural network, if one of the feature quantities of the input data is too sensitive, the neurons can be exchanged by the steps shown in FIG. Step 1) A sensor neuron having a standard linear input characteristic (f) during network formation. Step 2) Select and disconnect the neuron you want to change. Step 3) An object of a sensor neuron having an input / output characteristic (f) of an appropriate function is generated from a class (mold) library. Step 4) Link the newly generated neuron instead of the separated neuron. This operation has the same effect as filtering one of the features that is the input data on the spot.

Similar to neurons, synapses can be dynamically generated and deleted. For example, when there is theoretically no causal relationship between the feature amount which is a certain input data and the response neuron, or when the feature amount is erroneously recognized, the operation shown in FIG. 6 can be performed. Step 1) It is assumed that the sensor neuron and the response neuron are completely connected when the network is formed. In this case, all feature quantities are input to the response neuron. Step 2) Select unnecessary synapses and separate them. Step 3) The network structure is irregular, but works without any problems. This operation corresponds to narrowing down the feature amount. Also, if some evaluation amount can be set, the neural network itself can change the network structure during learning.

In this neural network, data is transmitted between neurons via synapses, so that the transmission path (that is, the network structure) can be changed simply by adding or deleting synapses. The separation of unnecessary synapses reduces the number of calculations in the learning process, shortens the calculation time, and reduces the parameters for calculation, which facilitates the convergence to the optimum solution.

In the present invention, as a method of inputting data to the neurocomputer, a method of using a feature quantity vector is used, unlike the method of using a pattern vector described above. This method is a method of extracting some feature amount from a pattern in a preprocessing stage and inputting the feature amount as a vector to a neural network. In this method, the process of extracting the feature amount requires a deep understanding of the target phenomenon, and since it takes time and labor, the weight of preprocessing is high and the load on the neurocomputer is low, so there is a possibility that learning will succeed. It is considered expensive. When using the feature vector, observe the causal relationship between the input feature and the internal state of the neurocomputer, and the output as described above, interactively analyze the inside, and exchange, connect, and disconnect neurons in the network. This can greatly improve the learning efficiency.

7 and 8 are explanatory views of an application system incorporating the neurocomputer of the present invention.

The neurocomputer system 8 comprises a preprocessor 2 for receiving the collected data 9 or the data collecting device 10 and data 1 such as flame temperature for controlling the combustion furnace, for example. The preprocessor 2 extracts a plurality of feature quantities from the received sensor data and inputs them into the database 3. Further, the characteristic amount stored in the database 3 of the neuro computer system 8 is input to the neuro computer 4 which simulates a neural network.

The output signal 5 of the neurocomputer is input to the output device 6 and converted into a control signal for controlling the equipment or plant 7. In a device such as a combustion furnace or a plant 7 to which a control signal is input, operation and operating conditions are changed by the control signal, and in the case of a combustion furnace, for example, in the case of a combustion furnace, the injection amount of fuel and the air amount are changed. Etc. are controlled and optimum operation is performed. The operating and operating conditions of the equipment or plant 7 are monitored by sensor signals such as the temperature of the combustion furnace, and fed back to the neuro computer system 8 as the data collection device 10 or the collected data 9.

The neurocomputer system 8 shown in FIG. 8 has a structure in which the preprocessor 2 and the neurocomputer 4 shown in FIG. 7 are integrated. Both the preprocessor 2 and the neurocomputer 4 operate with software on a Neumann type computer, and can be integrated.

A combustion furnace using a neurocomputer according to an embodiment of the present invention will be briefly described. This combustion furnace incinerates the garbage discharged day and night,
Due to problems such as pollution, it is necessary to control the concentration of exhaust gas such as CO2 and NOx. However, the types of garbage that are discharged day and night vary widely, and the operating conditions of the furnace, such as the amount of air and the amount of fuel that is injected, indicate the quality and quantity of the garbage that is input, and the combustion state such as the temperature of the flame inside the furnace. Must be changed from moment to moment depending on the situation.

Therefore, in the neurocomputer, for example, the brightness of the combustion flame in the furnace is read by the sensor, and the feature amount is extracted by the preprocessor to be used as the input data. The output data is the opening degree of the valve that controls the air amount,
As the teacher signal, the optimum operating condition is learned by using the exhaust gas concentration itself. After learning, the feature quantity is analyzed from the sensor data such as the temperature of the combustion flame, and the feature quantity is input for the input of various kinds of dust. Outputs the control valve opening. This achieves optimum operation of the refuse combustion furnace.

[0034]

According to the neurocomputer of the present invention, the feature amount data is input to the neural network,
While observing the state of learning, it is possible to exchange an arbitrary neuron with another neuron having different input / output characteristics, or to connect or disconnect the arbitrary neurons. Therefore, it becomes possible to train a neurocomputer capable of performing complicated processing such as optimal operation of a plant, which has been difficult as a practical problem in the past, and to put an applied device including the neurocomputer and a system including the same into practical use. Is possible.

[Brief description of drawings]

FIG. 1 is a block diagram of a neurocomputer according to an embodiment of the present invention.

FIG. 2 is an explanatory diagram of synapse and neuron classes (molds) according to an embodiment of the present invention.

FIG. 3 is an explanatory diagram showing the construction of a neural network according to an embodiment of the present invention, where (A) shows the inheritance structure of a class (mold), (B) shows the generation of synapses and neurons, and C) shows the construction of a network by linking synapses and neurons.

FIG. 4 is an explanatory diagram showing driving of a neural network according to an embodiment of the present invention.

FIG. 5 is an explanatory view showing exchange of neurons according to an embodiment of the present invention.

FIG. 6 is an explanatory diagram showing addition and deletion of synapses according to an embodiment of the present invention.

FIG. 7 is an explanatory diagram of a system including a neurocomputer according to an embodiment of the present invention.

FIG. 8 is an explanatory diagram of a system including a neuro computer according to another embodiment of the present invention.

FIG. 9 is an explanatory diagram of an example of a hierarchical neural network configuration.

[Explanation of symbols]

 1 means for receiving data 2 preprocessor 3 database 4 simulator 5 output means 6 output device 7 equipment or plant 8 neurocomputer system 9 collected data 10 data collection device 11 neuron and synapse generation means 12 network construction means 13 drive means 14 neural network 15 Neuron exchange, connection, disconnection means

Claims (4)

[Claims] 1. A learning means for generating a neuron and a synapse, a means for connecting the generated neuron and a synapse to construct a neural network, and inputting a feature amount and a teacher signal to the neural network. A neurocomputer comprising drive means, comprising means for exchanging any neuron constituting the neural network with another neuron having different input / output characteristics, or for arbitrarily connecting or disconnecting the neurons. 2. The synapse includes at least data of connection relations with input / output neurons and load data, and the neurons are composed of sensor neurons and response neurons, and at least data of connection relations with the synapses and response neurons, respectively. 2. The neurocomputer according to claim 1, further comprising: a learning function program, wherein the learning function of the response neuron rewrites the synaptic weight data. 3. A means for receiving data, a means for extracting a plurality of feature quantities from the data and inputting the feature quantity into a database, and a means for inputting the plurality of feature quantities stored in the database into a neural network. Neurocomputer application equipment. 4. The neurocomputer-applied device according to claim 3, an output device for converting an output signal of the neurocomputer-applied device, and a device for changing driving and operating conditions by the control signal of the output device. A mechanical device comprising: