JPH02268364A - Neural net - Google Patents

Abstract

PURPOSE:To quickly discriminate the permutation of a signal levels by dividing artificial neurons to hierarchies and giving differences in easiness of firing to them. CONSTITUTION:Artificial neurons 11 are two-dimensionally arranged to constitute a neural net. A bias signal J1 whose level is higher than the other bias signals is supplied to neurons 11 in a first row. Therefore, the output in neurons 11 in the first row is apt to be large. That is, neurons 11 in the first row fire most easily. Neurons 11 in second, third, and following rows fire less easily in order. When signals I1 to IM are supplied to neurons 11 from an input signal supply element 8, neurons 11 in the first row fire. That is, the input signal having a maximum level is discriminated. The output of these neurons 11 is supplied to the other neurons 11 to suppress firing these neurons 11. Next, neurons 11 to which the signal having the second highest level is supplied out of neurons 11 in a second row fire. Hereafter, neurons fire in the same manner.

Description

[Detailed Description of the Invention] [Object of the Invention] (Industrial Application Field) The present invention relates to the improvement of neural nets, for example, determining the magnitude order of a large number of similar signals based on their signal levels, This field relates to neural nets used to monitor and control plants and to determine the production order of product lots on production lines.

(Conventional technology) Monitors distribution systems such as gas turbine combustion temperature, boiler evaporator tube temperature, fuel cell cell stack temperature, fuel cell reformer temperature, heating furnace temperature, chemical reaction tank temperature, room temperature, and water pipe network pressure. When controlling or controlling a large number of similar analog input signals, the maximum value, intermediate value, minimum value, average value excluding the maximum and minimum values, average value of several values near the intermediate value, etc. of these signals can be calculated. A case can be considered in which a controlled object is controlled by using one operating end as a controlled variable.

This control method requires a magnitude ranking determination process to find the maximum value, minimum value, intermediate value, etc. of the signal levels (eg, current value, voltage value) of a large number of analog input signals. This signal processing is usually performed by repeatedly comparing the signal levels of two input signals.

In addition, we determine the order of patrols to visit multiple cities in the shortest distance, and determine the production order of product lots that minimizes the setup time when multiple product lots are produced on one production line. is called an NP-complete problem, and is determined by calculating evaluation index values for all combinations.

Recently, the use of neural networks for this type of decision has been considered (for example, l1opfleld, Tank
Co-authored with “NCural” Computation
or Decisions in Optimizati
on Problem, Biol, Cyber
No. 52.

1985, pp. 141 to 152) (Problems to be Solved by the Invention) In the conventional magnitude order determination method, iterative processing is necessary and time consuming in order to determine the magnitude order of the signal levels of input signals. . Therefore, if conventional signal processing is used,
Feedback - adds extra delay in the control loop,
Controllability cannot be improved.

Particularly when high-speed control is required, this delay in signal processing cannot be tolerated. It is also conceivable to select one of a large number of analog input signals and use it as a control variable. However, if the sensor that outputs the selected analog input signal becomes abnormal or the selected analog input signal fluctuates differently from other analog input signals, the entire distribution system will be controlled based on the selected analog human input signal, which will cause problems. There is a risk of disturbing the distribution system.

When performing upper limit checks and lower limit checks on individual input signals in υ signal monitoring, intermediate level signals that do not originally require upper limit checks or lower limit checks are also checked. Such a check would be time consuming or would make the alarm monitoring system bulky. Note that after determining the magnitude order, it is also possible to perform an upper limit check on the input signal of the maximum level and a lower limit check on the input signal of the minimum level. In this way, unnecessary checks become unnecessary. However, it takes time to check the order of magnitude, and there is a risk that the generation of an alarm will be delayed and a problem will occur.

In the conventional problem of determining a city tour route or determining the production order of variety lots, it is necessary to calculate evaluation index values for all combinations and select the optimal one, which takes time. In particular, as the number of cities to be visited and the number of 850 tons of products to be produced increase, the time required for calculation increases at a rate of about the cube of the increased number.
If the number of cities or lots is several dozen, it is virtually impossible to find a solution even using today's ultra-high-speed computers. For this reason, the use of neural networks to determine such solutions is being considered. Neural nets can be used to find solutions quickly. However, conventional systems using neural nets often encounter problems of this kind, such as when a certain city must be visited by a certain date and time, or when the evaluation index value cannot be calculated under constraint conditions such as delivery deadlines. The configuration is not designed to find the optimal solution.

This invention was made by focusing on the parallel processing performance of interconnected neural networks, and aims to provide a neural network with high application value and a system using the same.

Another object of the present invention is to provide a neural network and its use suitable for determining the order of signal levels at high speed or for determining the optimal ranking at high speed under constraint conditions such as deadlines. The objective is to provide a system that

(Means and effects for solving the problem) In order to achieve the above object, the neural
A net (network) is a two-dimensional array of artificial neurons based on threshold theory, interconnecting their inputs and outputs, and
We divided the artificial neurons into layers, and placed artificial neurons with different ease of firing in each layer.

Under such a configuration, the magnitude order of the signal levels of input signals is determined as follows. Artificial neurons are arranged in a matrix, and each artificial neuron is inhibited and coupled to each other so as to send a signal that suppresses the firing of other artificial neurons in the same column and row. Place the most firing artificial neurons in the first row, then the second, third, and so on.
......Sequentially, artificial neurons that are difficult to fire are placed. A signal to determine the order is input to each row of artificial neurons in the 221-ral net, and the neural net is operated. After a predetermined period of time, a stable state is obtained in which only one artificial neuron fires in each row and column. That is, among the artificial neurons that make up this neural net, the artificial neuron in the first row fires in the column to which the input signal of the highest level is supplied, and the other artificial neurons belonging to the first row and the same column fire. suppress. Next, the artificial neurons in the second row fire in the column to which the signal with the second highest input level is supplied, suppressing the firing of other artificial neurons belonging to the second row and the same column. In this way, the input signal to the column to which the artificial neuron firing in the first row belongs is 1
Indicates that the input signal has the th level. In this way, the order of input signals based on their signal levels is determined.

The order of cities to be visited and the production order of variety lots are determined as follows. Artificial neurons are arranged in a matrix, and rows are associated with cities and product lots, and columns are associated with ranks. At this time, those with earlier deadlines are assigned to rows with lower numbers. The combinatorial connections of the artificial neurons and the strength of the input are determined by correlating the energy function of the neural network with the evaluation index.

The artificial neurons in the i-th row and ttEj column are arranged such that the smaller the value of (i+j), the more likely they are to fire. The order of minimizing evaluation indicators such as total travel distance, total time required, total cost, and time for intelligent changeover is determined by the position in the matrix of the firing artificial neurons. for example,
In the first column, the city corresponding to the row to which the firing artificial neuron belongs is first visited or the variety lot is produced first, and the next city to visit or the variety lot to be produced is fired in the second column. Indicated by the line to which the artificial neuron belongs.

In such a configuration, by dividing the artificial neurons into hierarchies and varying the ease with which they fire, cities that you want to visit quickly or product lots that you want to produce quickly can be associated with rows with lower row numbers, thereby improving deadlines and evaluation indicators. It is possible to determine the ranking taking into account the optimization of

If a deadline must be strictly met, the neural pot network checks whether the deadline of each event is met in the determined order. If there is an overdue event, corrections are made such as changing the correspondence of the row corresponding to the event to a row with a lower row number, or making it easier for artificial neurons in columns with higher ranks to fire in the corresponding row. Thereafter, 2, the secondary net is operated again to determine the ranking that will not cause a deadline delay.

(Example) Hereinafter, a neural net (network) according to an example of the present invention and a system using the network will be described with reference to the drawings.

First, the artificial neuron used in this example will be explained with reference to FIG. The artificial neuron in FIG. 7 receives an external input signal 11, a bias signal J1, and a signal obtained by multiplying the output xi of another artificial neuron by a synaptic coupling coefficient ωl.

The artificial neuron adds input signals and generates an internal state u (t) according to the dynamics of the following equation (1). Note that in the case of inhibitory connections, the synaptic connection coefficient ωi is a negative value. However, if it is physically impossible to achieve a negative value,
The synaptic coupling coefficient ωi itself is a positive value, and is equivalently realized using the inverted signal xi (--xi) of the output xi of the artificial neuron.

+I+J ... (1) τ is the operating time constant.

The artificial neuron has this u (t)
is converted by an output function f to generate an output x (t).

z (t) = f [u (t), hl The output function is, for example, a sigmoid function as shown in FIG. 8, and if the threshold value is not included in the right-hand side as in equation (1), The threshold value becomes a parameter of the output function f.

Therefore, the ease with which this artificial neuron fires can be changed by changing the bias signal J, the threshold value, the shape of the output function f (see Figure 8), and the value of the synaptic connection coefficient ω.

The artificial neuron can be realized using, for example, an analog electronic circuit amplifier, or an optical system using a spatial modulation element. It is also possible to convert the above equation (1) into a difference equation and construct an artificial neuron using a digital arithmetic circuit.

Next, a neural network according to an embodiment of the present invention will be explained with reference to FIG. This neural net is a circuit that ranks input signals based on signal levels.

The neural net shown in FIG. 1 is constructed by arranging the artificial neurons 11 shown in FIG. 7 in a two-dimensional (matrix) manner. Each artificial neuron 11 is mutually inhibited and connected to other artificial neurons in the same row and to other artificial neurons in the same column. An input signal supply element (circuit, element) 8 is connected to the neural net 1 . The input signal supply element 8 supplies a plurality of input signals 11, I2, . A bias signal supply element 9 is also connected to the neural net 1 . Bias signal supply element 9 receives bias signal Jl
, J2...JM are respectively supplied to the artificial neurons 11 in the corresponding rows. Bias signals Jl, J2...
...J)l has a decreasing signal level in that order.

The output signal of each artificial neuron 11 is fed to a decision element 10.

Next, the operation of the circuit shown in FIG. 1 will be explained.

The artificial neurons 11 in the first row are supplied with a bias signal J1 having a higher level than the artificial neurons 11 in the other rows. Therefore, in the first row of artificial neurons, (1
) is large, u (t) tends to be large,
Therefore, the output x (t) tends to become large. That is, the artificial neurons in the first row are the most likely to fire. Below, the second
The artificial neurons 11 become more difficult to fire in the order of row, third row, and so on. In other words, in FIG. 1, the artificial neurons 11 arranged in a matrix are theoretically divided into layers in units of rows, and the artificial neurons 11 in the lower numbered layers are configured to fire more easily.

Signals It, I2, I3... from the input signal supply element 8°
...When the IM is supplied to the artificial neurons 11, the artificial neurons 11 in the first row to which the signal with the highest signal level is supplied is (1) among the artificial neurons 11 constituting this neural net 1. The value of I+J on the right side of the equation becomes the largest and fires. That is, the input signal with the output x=1 and the maximum level is determined. The output of the fired artificial neuron 11 is sent to other artificial neurons 11 in the same row and column. Since the artificial neuron 11 is inhibited, the synaptic retention coefficient ω is a negative value, and for the artificial neuron 11 in the same column and row as the fired artificial neuron 11, the second term Σω1xi on the right side of equation (1)
The value of (t) becomes smaller, ignition is suppressed, and stability is achieved.

Next, among the artificial neurons 11 in the second row, the artificial neuron 11 is supplied with the signal having the second highest signal level.
The value of I+J increases and fire occurs.

As a result, the signal having the second highest signal level is determined. The output of the fired artificial neuron 11 is the same -
The signal is supplied to the other artificial neurons 11 in the column and the second row, and the firing of these artificial neurons 11 is suppressed. After a certain period of time has elapsed since the input signal was supplied, a stable state is reached in which one artificial neuron fires in each row and column. The judgment element 10 receives the output signal of the artificial neuron 11, checks the signal level of the input signal row by row, and determines the input signal 11 to I.
Determine the order of signal levels of M. For example, when the artificial neuron in the first column fires in the first row and the artificial neuron in the second column fires in the second row, the determination element 10 determines that the level of the input signal is If, 12, etc. Outputs that it is in order. In this way, the magnitude order of the signal levels of the input signals can be determined within a short time.

Note that the input signal supply element 8 is composed of, for example, a plurality of variable current sources that output a current corresponding to an input signal.

Further, the determining element 10 is composed of, for example, a one-chip computer.

FIG. 2 shows an example of a specific configuration of the neural net 1 shown in FIG. 1. FIG. 2 shows two rows and six columns of artificial neurons arranged in a matrix and their connection relationships. ``In the circuit shown in Figure 2, the artificial neuron is composed of an analog amplifier. Further, synapses are composed of resistances, and by changing the resistance value, the degree of synaptic connection ω is changed. The inverted output of each amplifier is supplied to the amplifiers in the same column and row via resistors to achieve mutual suppression coupling.

Next, an embodiment in which a plant is actually controlled using the circuit shown in FIG. 1 will be described with reference to FIG. 3.

This example shows an example of controlling the temperature of a gas turbine. A plurality of burners 21 are attached to the gas turbine. Combustion fuel is supplied to the burner 21 via a common fuel vibe 23. The total amount of fuel supplied to the burner 21 is adjusted by the degree of opening and closing of the valve 22.

A plurality of temperature sensors SE are arranged within the gas turbine,
The output signal of each sensor is supplied to a signal processing circuit 24. The signal processing circuit 24 has, for example, the configuration shown in FIG. 1 and an average level calculation circuit. This average level calculation circuit receives the output signal of sensor SE and the output signal of judgment element 10. The average level calculation circuit calculates, based on the judgment result of the judgment element 10, the average value of the signal levels of signals other than the signal having the maximum signal level and the minimum signal level, or the average value of the signal levels of a plurality of signals near the intermediate value. A signal indicating the value is output to the control circuit 25. This signal indicates the average temperature inside the gas turbine. The control circuit 25 is the signal processing circuit 2
In response to the signal indicating the average value from 4, a control operation (for example, a PID operation) is performed to adjust the opening/closing degree of the valve 22 and control the temperature inside the gas turbine.

In this embodiment, the temperature inside the gas turbine is not constant, but varies depending on the position, resulting in a distributed system.

For this reason, a plurality of temperature sensors SE are arranged within the gas turbine to determine the average value of the combustion temperature.

On the other hand, the temperature inside the gas turbine is extremely high, and there is a high probability that the sensor will fail. Therefore, in this embodiment, the average temperature is determined using only signals with a high probability of being normal signals among the output signals of the sensor SE. Therefore, the reliability of the control is high, and since the signal processing circuit 24 is a circuit using the neural network shown in FIG. 1, high-speed operation is possible, and online control is possible.

In this way, according to this embodiment, by utilizing the parallel processing ability of the neural network, the order of a large number of input signals can be determined almost instantaneously, based on their signal levels, without repetitive operations. Therefore, signals having the maximum signal level, minimum signal level, intermediate signal level, etc. can be selected in a short time.

As a result, it is possible to obtain a highly reliable average value excluding signals with a high possibility of input abnormality, or a highly reliable average value using several input signals near the intermediate value, with a short detection delay time. becomes easier. By using the obtained average value as the control amount, the distribution system can be controlled favorably and with high reliability.

In addition, instead of performing upper limit and lower limit checks for many input signals one by one, only the signal with the maximum signal level is checked as the upper limit, and only the signal with the minimum signal level is checked as the lower limit. Upper and lower limit alarm processing can be easily performed.

Next, a second embodiment of the invention will be described with reference to FIG. The neural net 2 of this embodiment is constructed by arranging the artificial neurons 12 shown in FIG. 7 in two dimensions (in a matrix). In this embodiment, the signal levels of the bias signals supplied to the artificial neuron 12 are all zero or constant values. Each artificial neuron 12 is mutually inhibitedly coupled to other artificial neurons 12 in the same row and column as the one to which it belongs. Furthermore, in this neural net 1, the mutual inhibition coefficient ω1 regarding the input of the artificial neuron 12 in the first row is the mutual inhibition coefficient ω2 regarding the input of the artificial neuron 12 in the other row.
It is the largest compared to ~ωM (negative number, smallest absolute value, weakest suppressive force). Below, 12 lines, 3rd line...
. . . The mutual inhibition coefficients ω2 to ωH in the Mth row become smaller in order. Furthermore, neural net 2 has the 15th
The same input signal supply element 8 as in the embodiment is connected. Although not shown, the determination element 10 is also
Connected to Net 2. Two knees configured like this
In the ral net, if we consider that the state of xi (t) in equation (1-) is almost the same in all rows, then the second term on the right side Σω1. The value of xl(t) is the largest for the artificial neuron in the first row (it is a negative number and has the smallest absolute value). Therefore, the first
The inhibitory power of the artificial neurons in the row is the weakest, and they are more likely to fire. Thereafter, the firing becomes more difficult in the order of the second line, the third line, and so on. Therefore, by the same effect as in the embodiment of FIG. 1, the magnitude order of the signal levels of the input signals can be determined in a short time.

As described in conjunction with FIG. 4, the same effect can be obtained by setting the threshold value shown in FIG. can be realized.

The same thing can also be achieved by making the output function of the artificial neuron different for each row.

In this case, the output function f should be set to the one that is most likely to fire for the artificial neurons in the first row (for example, the solid line in Figure 8).
, the second line is made to be harder to fire than the first line (for example, the one-dot chain line in FIG. 8), and thereafter the third, fourth, and so on lines are made to be harder to fire.

Next, a third embodiment of the present invention will be described with reference to FIG.

The neural net 3 of this third embodiment is constructed by arranging the artificial neurons 13 shown in FIG. 7 in a matrix and connecting them to each other (not necessarily suppressing connections). The artificial neuron in the i-th row and first column of neural net 3 is (
Layer 1J that is the same as the artificial neuron with the same value of i+j)
They are hierarchically divided according to their belonging. Artificial neurons belonging to the same hierarchy have the same tendency to fire. (i+j
), for example, by adjusting the threshold h1j, artificial neurons that are more likely to fire are arranged as base layers. Input signal Ilj and bias signal J1
An input/bias signal supply element 4 supplying j is connected to the input terminal of each artificial neuron 13. A checking element 5 that checks whether the solution generated by the neural net 3 is a valid solution is connected to the output terminal of each artificial neuron 13.

A parameter setting element 6 for changing (correcting) the parameters of the neural net 3 such as input signal 11 bias signal J1 coupling coefficient ω, threshold value, etc. when the solution generated by the neural net 3 is an unacceptable solution is configured for each artificial neuron and It is connected to the input/bias signal supply element 4. A control element 7 that manages and controls the operations of these elements is connected to the neural net 3, the check element 5, and the parameter setting element 6.

The operation of the neural net system shown in FIG. 5 will be explained. In order to make it easier to understand, an example will be explained in which the production order of lots of different types is determined so as to reduce the cost of intelligently changing lots of different types on a production line.

First, each column of the neural net 3 is associated with a production rank. Next, each row is associated with a product lot. In this case, the earlier the product lot is produced, the lower the row number is associated with it. In other words, the firing of the artificial neuron in the i-th row and first column means that the variety lot associated with the i-th row will be produced in the j-th manner.

The output of the artificial neuron, which means to produce variety lot The solution that can be claimed especially for the number of species rods is the one that minimizes the following energy function E (second equation).

E-(A/2)・MojihdaykVX/ (VY, +1 + VY, k() ... (2) A, B, C, and D are constants.

Here, the first term on the right side becomes O when only one or less artificial neurons fire in each row, the second term on the right side becomes 0 when only one or less artificial neurons fire in each column, and the third term on the right side becomes O. is 0 when only n artificial neurons fire.
The fourth term means the total switching setup cost.

On the other hand, a mutually coupled neural network has the property of being stable in a state where the following energy function E' (Equation 3) becomes minimum.

1 Σ (I Xk + J , ωX k , Yi key -A - B conflict - D - dXY (1-δ to δ" (1-dXY) - C dXY (δ, P, +t +6 years old, to -1>δ/ −
1: −J δ, P−0: fJ The input signal I Xk (bias human power JXk is zero) of the artificial neuron in the The third term on the right-hand side of is extremely small, so if we ignore it, we get E-
E”.

I Xk=c n (5) When the neural net with the coupling coefficients and input signal values determined in this way is operated, the energy function value becomes a stable state with a minimum value. The production order that minimizes the total switching setup cost is determined based on the firing state of the artificial neurons at that time.

For example, in a stable state, artificial neurons in 1 row and 1 column, artificial neurons in 2 rows and 3 columns, artificial neurons in 3 rows and 2 columns...
... is firing, this neural net instructs that product lots A, C, B, etc. should be produced in that order.

In this embodiment, the threshold 1iihlj of the artificial neuron 13 in the i-th row and j-th column is made different by a positive correlation with the value of (1+J), and the ease with which the artificial neuron 13 fires is changed for each layer. For this reason, a diagonal line downward to the right in the drawing (a line connecting the artificial neuron in row 1 and column 1 and the artificial neuron in row M and column M)
Artificial neurons placed near are likely to fire. That is, it is set that product lots associated with younger rows are likely to have an early production order because of their short delivery times. Therefore, in this embodiment, it is possible to obtain a production order solution that compromises the minimum total changeover setup cost and the delivery date.

In the above explanation, we talked about changing the ease with which artificial neurons fire by changing the threshold.
This can also be realized by changing the shape of the output function f shown in FIG. 8 or by changing the value of the bias input signal J. Furthermore, it can also be realized by adding a different bias for each layer to the coupling coefficient ω determined from the above-mentioned energy function and correcting the coupling coefficient.

Next, check element 5 will be explained. In Neural Middle Net 3 described above, rankings are determined in consideration of delivery dates, but if there is a variety lot for which delivery dates must be strictly adhered to, there is no compensation for determining the rankings to ensure that the delivery dates are met. Therefore, the check element 5 checks whether the production of each product lot is completed by the respective production deadline based on the following equation (6).

In the sixth equation, Tα(1-1,1) means the time required to switch from the i-th, first product type to the first product type. Tα(0,1) means the initial setup time. Tβ(,1) means the time required to produce the first variety to be produced. Tγ(1) means the production deadline of the first variety to be produced. The parameter setting element 6 receives the check result of the check element 5. If there is a product lot that will be overdue, the parameter setting element 6 transfers the row corresponding to the product lot to a younger row.

The parameter setting element 6 also changes the coupling coefficient ω of the neural net 3 based on equation (4) as the rows are changed.

For example, specifically, in the order of items FJiASCSB...
Let us consider the case where the solution to produce is obtained. For example, if it is determined that the delivery date of item t8Ii B cannot be met by checking check element 5, parameter setting element 6
transfers type B to the first line and type A to the second line, and changes the coupling coefficient accordingly. When the transfer is completed, the control element 7 starts up the neural net 3 again,
Find a new solution.

The control element 7 operates on each element to automatically repeat the above-mentioned operations until a solution is obtained that satisfies the delivery date for all products. The operation of the control element 7 will be explained with reference to the flowchart in FIG. First, the control unit 7 instructs the parameter setting element 6 to set initial parameters for the neural net 3 and the input/bias signal supply element 4 (step Sl). When the initial parameter settings are completed, the neural net 3 is activated (step S2
). The control element 7 checks whether the neural net 3 has reached a stable state, for example by measuring time (step S3). When the neural reference network 3 reaches a stable state, the control element 7 instructs the check element 5 to check the deadline (step S5). When a lot of a variety that cannot meet the deadline is found through the operation of the check element 5 (step S5), the control unit 7 resets the parameters of the neural net 3 so that the production order of the lot of that variety is improved. is commanded to the parameter setting element 6 (step S6). After completing the settings (step S7),
The control unit 7 restarts the neural net 3 (step S2). The control element 7 continues this series of operations until it is determined in step S5 that there is no overdue product lot.
Have them repeat.

In this way, depending on the firing state of the artificial neuron, it is possible to obtain the optimal production order in a short time under conditions with deadline constraints.

In addition, if a solution is obtained in which there is a product with a late deadline,
Without changing the rows, the obtained rank of the row to which the product lot is assigned is supplied to the artificial neurons 13 in such a way that the artificial neurons belonging to columns younger than the column to which the lot is assigned are more likely to fire. A similar effect can be obtained by correcting the signal level, threshold value, or output function of the bias signal. For example, if the product g1f3 assigned to the second row is produced third and a solution that causes the deadline to be delayed is obtained, a bias signal with a large signal level is sent to the artificial neurons in the first and second columns of the second row. It may be supplied. Alternatively, the threshold values of the artificial neurons in the first and second columns of the second row may be reduced, or the output function f may be corrected.

If the above-mentioned quality lot name is replaced with a city name and the production order is replaced with a traveling order, the embodiment of the m5 diagram can be applied to the case of finding a solution to the traveling salesman problem. Moreover, it is not limited to these, and can be applied to answers to other similar questions.

In addition, input/bias signal supply element 4, check element 5
, the parameter setting element 6, the control element 7, etc., can be realized individually or integrally using a microcomputer.

[Effects of the Invention] According to the present invention, by utilizing the parallel processing ability of the neural net, a large number of input signals are processed according to their signal levels.
The ranking can be determined almost instantaneously without repeated operations. Therefore, it is possible to obtain a neural net that facilitates selecting signals having the maximum signal level, minimum signal level, intermediate signal level, etc. in a short time. As a result, it is possible to obtain a highly reliable average value excluding signals with a high possibility of input abnormality, or a highly reliable average value using several input signals near the intermediate value, with a small detection delay time. becomes easier. By using the obtained average value as the control amount, the distribution system can be controlled favorably and with high confidence.

Also, instead of checking a large number of input signals one by one to see if they exceed the upper limit or below the lower limit, only the signal with the maximum signal level is checked as the upper limit, and only the signal with the minimum signal level is checked as the lower limit. By checking this, you can easily perform upper and lower limit alarm processing for the distribution system.

Furthermore, it is possible to obtain a neural net and a system that can quickly determine the optimal visit order under conditions where there are 171 appointments on the visit date and the optimal production order of product lots under delivery date constraints.

[Brief explanation of drawings]

FIG. 1 is a configuration diagram of an embodiment of a mutually inhibiting coupled neural net system in which the bias signal value is changed for each layer to change the ease of firing in order to determine the magnitude order of the signal level of the input signal. Figure 2 shows the neural network shown in Figure 1.
A circuit diagram showing an example of a specific configuration of a part of the net, Fig. 3 is a block diagram showing an example of a control system using the neural net system shown in Fig. 1, and Fig. 4 shows a signal level of an input signal. Figure 5 is a block diagram showing the configuration of a mutually inhibiting coupled neural net system in which the ease of firing is changed by changing the coupling coefficient value or threshold value for each layer in order to determine the magnitude order of . A block diagram showing the configuration of a mutually coupled neural net system that can determine the optimal ranking in FIG.
FIG. 7 is a schematic diagram of the artificial neuron, and FIG. 8 is a diagram showing the form of the output function of the artificial neuron. 1.2.3...Neural net, 4...Input
Bias signal supply element, 5... Check element, 6...
- Parameter setting element, 7... Control element, 8... Input signal supply element, 9... Bias signal supply element, 10
... size IfrW element, 11.12.13... artificial neuron. Applicant's representative Patent attorney Takehiko Suzue Figure ω; 1st line Shichoasya 1!1 book song (ω1〉 ω2〉 ...〉ωM) hself: '$ i Kdingσ) Shiki- 11a (h'<h2<...<hM) Diagram Tα(i-1,i) Tβ(i); No. 1-141! ± Two parts, one each! ! 1 ko 1 thick T-fila product ↑ 1 raw N door! q〒54
゜Tγ(1): Fuikome (；ヱ1Troaon's 1st voice period Tora. Figure Figure Figure

Claims (1)

[Claims] (1) In a neural network in which interconnected artificial neurons are arranged in a two-dimensional manner and parallel information processing is performed,
A neural net characterized by dividing artificial neurons into layers and using artificial neurons that have different firing susceptibilities for each layer. (2) The neural net according to claim 1, wherein the artificial neurons are arranged in a matrix and divided into layers by row or column. (3) A claim characterized in that the artificial neurons are arranged in a matrix, and the artificial neurons in the i-th row and j-th column belong to the same layer as the artificial neurons with the same (i+j) value. The neural net described in 1. (4) The neural net according to claim 1, wherein the artificial neurons in the same row or column have mutual inhibitory connections. (5) The neural network according to claim 1, wherein the artificial neurons are interconnected so that a stable point of the state of the neural network exists in a state where a predetermined energy function becomes a minimum. (8) The neural network according to claim 1, wherein the artificial neurons are arranged in a matrix, the artificial neurons in the same row or column are mutually inhibited, and each row or column is divided hierarchically. . (7) Arrange the artificial neurons in a matrix, connect each artificial neuron to each other so that the stable point of the neural net state is the state where the predetermined energy function is minimum, and 2. The neural net according to claim 1, wherein the artificial neurons are hierarchically divided so that the artificial neurons having the same value of (i+j) belong to the same hierarchy. (8) The neural network according to claim 1, wherein the ease of firing of each layer of the artificial neuron is varied by changing the magnitude of a bias signal applied to the artificial neuron for each layer. . (9) The neural net according to claim 1, wherein the ease with which the artificial neurons fire for each layer is made different by changing the threshold value of the artificial neurons for each layer. (10) The neural net according to claim 1, wherein the ease with which the artificial neurons fire for each layer is varied by changing the shape of the output function of the artificial neurons for each layer. (11) The neural net according to claim 1, wherein the ease with which the artificial neurons fire in each layer is made different by changing the strength of the inhibitory connections of the artificial neurons in each layer. (12) Claim 4 characterized in that the artificial neurons are hierarchically divided by row or column, and the artificial neurons belonging to the same row or column are provided with an input signal supply element that supplies the same input signal. Neural net described in. (13) The feature is that the row or column of the artificial neurons is associated with a predetermined event, and the column or row of the artificial neuron is associated with the order of occurrence of the predetermined event, and is used for determining the order of occurrence of the predetermined event. 8. The neural net according to claim 7. (14) A check element that checks whether there is an event that is too slow in the ranking determined by the neural network, and a check element that checks whether or not there is an event that is too slow in the ranking determined by the neural network, and if there is an event that is too slow, the ease with which artificial neurons related to that event fire. 14. The method of claim 13, further comprising: a parameter setting element for modifying; and a control element for controlling the operation of the repeating, neural net and checking element and the parameter setting element until a desired result is obtained.
Neural net for ranking determination described in . (15) A check element that checks whether there is an event that is too slow in the order determined by the neural network, and if there is an event that is too slow, the row corresponding to that event is set to a row with a lower number than that row. and a parameter setting element that changes the strength of mutual connections between artificial neurons as the changes occur, and a control that repeatedly controls the operation of the neural net, check element, and parameter setting element until the desired result is obtained. Claim 13 characterized in that it comprises an element.
Neural net for ranking determination as described. (16) Determine the order of visiting destinations to minimize the visiting distance, total time required, and total cost under the constraint that the specified destinations be visited within a predetermined period as visiting destinations for visiting events. 14. The neural network for ranking determination according to claim 13. (17) Let the event be a variety lot produced on a production line,
14. The method according to claim 13, wherein the production order of product lots is determined to minimize the total cost and total time for switching product lots under the constraint that production of a predetermined product lot is completed within a predetermined date and time. Neural net for ranking.