JPH06195485A - Neural network outputting membership function - Google Patents

Abstract

PURPOSE:To make the type of an NN have flexibility and to obtain a membership function by a balanced arrangement, in the NN outputting the membership function. CONSTITUTION:This neural network(NN) has a cell 2 and a bias cell 1 outputting input signals, plural first intermediate cells 3 to 6 to be connected with the signals of the cells by a prescribed coupling coefficient, plural second cells 7 to 9 to be connected with these intermediate cells by the prescribed coupling coefficient and two output cells 10, 11. An educator signal of a fixed value is impressed via the one side 10 of the output cells.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a neural network (hereinafter referred to as NN) that outputs a membership function.

[0002]

2. Description of the Related Art In recent years, a neuro computer has been developed and various research results have been announced. The neurocomputer focuses on neurons (nerve cells), which are the basic elements of the brain, and tries to achieve the same function as the brain by using the NN resulting from the combination of these as a hint. The feature of these NNs lies in the parallel information processing between the neurons and the ability of learning. Moreover, the actual network model consists of an input layer, a middle layer, and an output layer, each of which has cells corresponding to each neuron, and cells in each layer are connected to all cells in the layers before and after that layer. ing. In addition, a coupling coefficient called a load is introduced for coupling each cell, and these coefficients are corrected by learning.

[0003]

As described above, according to the conventional technique, the cells belonging to each layer forming the NN are combined with all the cells in the layers before and after that. Therefore,
In the conventional hierarchical NN learning program, calculation is performed assuming that one cell is combined with all cells in the layers before and after it, and the NN type is inflexible and the output waveform has a partial shape. There is a drawback that it causes a deviation. The present invention has been made to solve the above drawbacks, and an object of the present invention is to provide an NN that gives flexibility to the NN and outputs a membership function with an even arrangement.

[0004]

To achieve the above object, the present invention provides a cell for outputting an input signal and a bias cell, and a plurality of first cells connected to the signal of the cell via a predetermined coupling coefficient. An intermediate cell, a plurality of second intermediate cells connected to these intermediate cells with a predetermined coupling coefficient, and two output cells, and a constant-value teacher signal is applied via one of the output cells. It was configured to do.

[0005]

Embodiments will be described below with reference to the drawings. Figure 1
FIG. 1 is a configuration diagram of an embodiment of a NN according to the present invention. In the figure, 1 to 11 are cells, of which cell 1 is a special cell, called a bias cell, which always outputs 1. W
₁ ~W _7, WB _{1 ~WB 4} is a coupling coefficient. Cell 3-
The input / output function of 6 is the sigmoid function shown in equation (2), and the input / output function of cells 7 to 11 is the identity function shown in equation (1). The coupling coefficients of cells 4 and 8 and cells 6 and 9 are fixed to -1 as shown in the figure, and learning is not performed. The cell 7 has the coupling coefficient 1 from the cell 3 and the bias cell 1
, The coupling coefficient 1 from the cell 6 and the coupling coefficient 1 from the bias cell 1 are also input to the cell 9. Further, the cell 10 is a cell for learning so that the total sum of the outputs of the cells 7 to 9 approaches 1. Therefore, the teacher signal 1 is given to the cell 10, and the cells 10 and 7
The coupling coefficient with 9 is fixed to 1 and not learned.

[Equation 1] y = x …………………………………… (1) y = 1−exp (−x) / 1 + exp (−x) ………… (2) However, x: sum of input to cell, y: output of cell. FIG. 3 shows the outputs of cells 7, 8 and 9 shown in FIG. Cell 7
For, the sigmoid function from cell 3 is input as it is (coupling coefficient 1), and the coupling coefficient from bias cell 1 shifts to the right. The output of cell 8 is output as the sum of the inverted output of the sigmoid function from the coupling coefficient (-1) from cell 4 and the sigmoid function itself from cell 5, and the output of cell 9 is the combination from cell 6. Due to the inverted output of the sigmoid function by the coefficient (−1) and the shift by the coupling coefficient 1 from the bias cell 1. The coupling coefficients W _{1 to} W ₄ determine the smoothness of the membership functions of the cells 7 to 9, and WB _{1 to} WB ₄ are the cells 7 to 9.
The X-axis offset of the membership function of 9 is determined.

FIG. 2 is a block diagram of another embodiment. The difference from FIG. 1 is that two cells are used for handling a sigmoid function from an NN (a symbol A in the figure) having a membership function function. , The connection between them is as shown in FIG. Similar to the above-described embodiment, the cell 8 receives the simple addition output from the cells 5 to 7 and gives the teacher signal 1 to the cells 8 and 5 to 5.
The coupling coefficient with 7 is fixed to 1 and is not learned. FIG. 4 is a membership function waveform by cells 5 to 7, and the creation thereof is the same as in the case of FIG. Also in this embodiment, the same effect as in FIG. 1 can be obtained.

[0007]

As described above, according to the present invention, since a teacher signal having a constant value is provided to the cell that outputs simple addition, learning can be performed so that the total sum of the membership function outputs is constant. , The membership function is obtained in a biased arrangement.

[Brief description of drawings]

FIG. 1 is a configuration diagram of an embodiment of a NN according to the present invention.

FIG. 2 is a configuration diagram of another embodiment.

3 is an output characteristic diagram of cells 7, 8 and 9 in FIG.

4 is an output characteristic diagram of cells 5, 6, and 7 of FIG.

[Explanation of symbols]

1 to 11 cells W _{1 to} W ₇ , WB _{1 to} WB ₄ coupling coefficient

Claims (1)

[Claims] 1. A cell that outputs an input signal and a bias cell, a plurality of first intermediate cells that connect the signals of the cells through a predetermined coupling coefficient, and these intermediate cells and a predetermined coupling coefficient. A neural network having a plurality of second intermediate cells connected to each other and two output cells, wherein a teacher signal having a constant value is applied through one of the output cells, and outputs a membership function. .