JPH01177603A - Construction for adapting control mechanism - Google Patents

Abstract

PURPOSE:To easily adapt a control mechanism by executing a self- systematization in a network so as to cause a difference between a layer network and an educator signal to be small. CONSTITUTION:The difference by comparing the output of a layer network 2 with a control input 1 given to a system and an educator signal 4 is fed back to the correcting functions of the weight of the layer network 2 and a bias. The self-systematization is executed in the network so that the difference between the layer network 2 and the educator signal 4 can be made small. Thus, the adaptation by a trial and error can be made unnecessary, and an adapting control mechanism can be obtained in which the knowledge and experience of a skilled worker to already exist can be effectively utilized.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention provides an adaptive control mechanism that changes its control capability in response to a teacher signal, particularly an adaptive control mechanism that has a layered network type structure in its structure. The present invention relates to the structure of an adaptive control mechanism that has an adaptive control mechanism.

[Conventional technology]

In conventional control mechanisms, whether PID control or fuzzy control, human experts create the information needed for control, adjust that information through trial and error, and adapt the control mechanism. Met.

[Problem that the invention seeks to solve]

However, this method requires a large amount of trial and error even for simple control, and there is a problem in that if various conditions change, the same trial and error must be performed again.

Furthermore, it is especially important that in such work,
The problem is that the ability of system experts is required, and the knowledge that already exists as the experience of the workers cannot be used effectively.

It is an object of the present invention to provide a structure for an adaptive control mechanism that eliminates the need for such trial-and-error adaptation and makes more effective use of the knowledge and experience of already existing skilled workers. be.

[Means for solving problems]

The present invention provides a structure of an adaptive control mechanism having one or more inputs and one or more outputs, which a) has one or more signal input functions, and b) has a signal input function for each of the inputs. C) and has a function to modify the weight; d) has a function to calculate the sum of products of the input and the weight; and e) adds a variable bias to the sum of products. f) and has a bias correction function; g) and
A node has a mapping function that maps the calculation result to one or more real numbers, h) and has one or more signal outputs that output the mapped value, and 1) the node has a plurality of J) has a network having a structure arranged in individual layers; J) and has one or more teacher signal inputs; k) and compares the teacher signal with the output of the network. 1) and an output of the comparison function is connected to the weight correction function and the bias correction function.

It is.

〔Example〕

FIG. 1 shows a schematic diagram of the configuration of an embodiment of the present invention. In the figure, number 1 indicates a control input, 2 indicates a layered network, 3 indicates a control output, 4 indicates a teacher signal input, and 5 indicates a comparison function between the teacher signal and the control output.

In this configuration, the difference between the output of the layered network due to the control input applied to the system and the teacher signal is fed back to the weight and bias correction function of the layered network, so that the difference between the layered network and the teacher signal is small. Self-organization takes place in the network so that

Figure 2 shows a conceptual diagram of a layered network.

The number 6 in the figure indicates one node constituting the network, and the number 7 indicates an arc=5- connecting the nodes.

FIG. 3 is a conceptual diagram of the functions of the nodes that make up the network. In the figure, number 8 is the variable weight of the input, 9 is the product sum function of the input with that weight, 10 is the variable bias, 11 is the addition function of the bias and product sum, and 12 is the result. Demonstrates the mapping function.

FIGS. 4 and 5 show examples of the operation of the adaptive control mechanism having this configuration. To simplify the example, two analog quantities will be considered as control inputs, and one analog quantity from 0 to 1 will be considered as control output.

Figure 4 shows input 1 and input 2 on the vertical axis as an example of the teacher signal.
The contour lines of the control output are shown on the horizontal axis. In the figure, number 13 indicates a contour line of control output 1, and number 14 indicates a contour line of control output 0.5. In this relationship, the layered network is trained using only eight pieces of data indicated by white circles 15 and 16, respectively, as examples of teacher signals.

FIG. 5 shows the results of learning the training data shown in FIG. 4 into a layered network. The figure shows input 1 and input 2 as in Figure 4.
The contour lines of the control output are shown when the vertical and horizontal axes are taken as the vertical and horizontal axes. In the figure, number 17 indicates a contour line of control output 1, 18 indicates a contour line of control output 0.5, and 19 indicates a contour line of control output 0.75.

As is clear from the figure, in the cases of control outputs 1 and 0.5, even though only four pieces of teacher data were provided, the intermediate values were also completely complemented. Further, as shown in 19, a value considered to be reasonable is given as the control output even in the case of the control output area 75 for which no teacher data was explicitly given.

In other words, it can be said that this system is adapted to this control system by using only eight pieces of teacher data.

This training data may be something like a mathematical formula if the operation is theoretically clear. Furthermore,
Even if such a theory is not clear, for example, if a skilled worker exists, the control mechanism can be developed by sampling the worker's work and using it as a teacher, or by simply providing a desired output for a certain input. It becomes possible for people to adapt to the system by learning.

〔Effect of the invention〕

As is clear from the above description, according to the present invention, it is not necessary to repeat a huge amount of trial and error in order to adapt the system to the situation in various control mechanisms (and furthermore, it does not require specialized knowledge to do so). It is possible to realize a structure that adapts the control mechanism by simply providing teacher or model data without requiring it.

[Brief explanation of the drawing]

FIG. 1 shows a schematic diagram of the configuration of an embodiment of the present invention. Number 1 in the figure is the control input, 2 is the layered network,
3 shows a control output, 4 shows a teacher signal input, and 5 shows a comparison function between the teacher signal and the control output. Figure 2 shows a conceptual diagram of a layered network. In the figure, number 6 indicates one node constituting the network, and number 7 indicates an arc connecting the nodes. FIG. 3 is a conceptual diagram of the functions of the nodes that make up the network. Number 8 in the figure is the variable input weight, 9 is the product-sum function of the input, 10 is the variable bias, and 1 is the variable input weight.
1 indicates the addition function of bias and sum of products, and 12 indicates the mapping function of the result. FIG. 4 and FIG. 5 are diagrams showing an example of the operation of the adaptive control mechanism having this configuration. Figure 4 shows input 1 and input 2 on the vertical axis as an example of the teacher signal.
A diagram showing contour lines of control output when the horizontal axis is used. Number 13 in the figure indicates a control output 10 contour line, and 14 indicates a control output 0.5 contour line. Furthermore, white circles 15 and 16 indicate data used as teacher signals in learning. FIG. 5 is a diagram showing the result of learning the teacher data of FIG. 4. In the figure, number 17 indicates a contour line for control output 1, 18 indicates a contour line for control outputs 0 and 5, and 19 indicates a contour line for control output 0.75. Applicant: Seiko Epson Co., Ltd. Figure 1 Figure 2 Figure 3 Enclosure h2 Figure 4 - Figure 5 - 1 -

Claims (1)

[Claims] A structure of an adaptive control mechanism having one or more inputs and one or more outputs, which a) has one or more signal input functions, and b) each of the inputs. c) and has a function to modify the weight; d) has a function to calculate the sum of products of the input and the weight; and e) has a variable bias to the sum of products. f) and has a bias correction function; g) and
A node has a mapping function that maps the calculation result to one or more real numbers, h) and has one or more signal outputs that output the mapped value, and i) has multiple nodes. j) has a network having a structure arranged in individual layers, j) and has one or more teacher signal inputs in its structure, and k) and compares the teacher signal with the output of the network. 1) and the output of the comparison function is connected to the weight correction function and the bias correction function.