JP3523325B2 - Neural network, signal processing device using the same, autonomous system, autonomous robot, and mobile system - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a neural network imitating a nerve cell, which can be applied to various controls such as recognition of images and voices, adaptive control of robots, associative memory, nonlinear prediction, etc. The present invention relates to a signal processing device, an autonomous system, an autonomous robot, and a mobile system used.

[0002]

2. Description of the Related Art The function of a nerve cell (neuron), which is a basic unit of information processing in a living body, is mimicked, and further, this "nerve cell mimicking element" (nerve cell unit) is networked to perform parallel processing of information. What we aimed for was a so-called neural network. Although it is easy to perform character recognition, associative memory, motion control, etc. in a living body, there are many things that conventional Neumann computers cannot easily achieve. Attempts have been actively made to solve these problems by imitating the neural system of a living body, in particular, the functions peculiar to the living body, that is, parallel processing, self-learning, etc. by a neural network.

However, most of the studies on these neural networks deal with static states, which are different from the dynamic states that are actually performed in the living body. Therefore, in order to realize the information processing function of humans and other living bodies, it is necessary to handle this dynamic state.

The "chaos phenomenon" is considered to play a major role in the information processing of dynamic neural networks. An example of the neural network model having this chaotic state is shown in "Chaos in Neural Systems" (edited by Kazuyuki Aihara, Tokyo Denki University Press, pp. 157-188).

[0005] Figure 2 2 shows a block diagram construction of an electronic circuit model of chaotic neurons has been shown in this document. This uses two sample-and-hold circuits (S & H circuits) 1 and 2 and a one-dimensional mapping circuit 3, and an S & H circuit 1 holds an internal state value one step before according to a clock 4, and this is stored in a one-dimensional mapping circuit. 3 and the value of the next time is obtained at the output of the S & H circuit 2 in accordance with the clock 5 and the operation is repeated.

[0006]

However, the chaotic neuron model as shown in the above document is composed of an analog circuit as shown in FIG. Furthermore, although it is a simple model, it still has complexity, such as the existence of a refractory period. Therefore, the function of the dynamic neural network is not yet fully utilized.

[0007]

A neural network according to a first aspect of the present invention comprises an adding means for adding a plurality of signals,
Non-linear conversion means for performing non-linear conversion of the internal activation value resulting from the addition of the addition means and outputting the conversion result to the outside as a neuron output value; And external input means for performing arithmetic processing on the internal activation value and the neuron output value to output to the adding means, and the displacement amount of the internal activation value is changed according to the neuron output value. (Sum of products of input value of external input signal and coupling coefficient value)
+ (Threshold value)-(product of internal activation value and neuron output value)
A plurality of neuron elements having multiple inputs and one output, which are calculated and determined as, are connected to each other.

According to a second aspect of the present invention, in addition to the structure of the first aspect of the neural network, a storage means is provided which is connected to the addition means and stores the internal activation value.

In the neural network according to a third aspect of the present invention, in the configuration of the neural network according to the second aspect, the storage means for storing the threshold value for determining the displacement amount of the internal activation value is also used as the storage means for the coupling coefficient. It is a thing.

According to a fourth aspect of the present invention, in addition to the structure of the second or third aspect of the neural network, latch means is provided on the output line from the external input means to the adding means.

According to a fifth aspect of the present invention, there is provided a signal processing device, wherein the neural network according to the first, second, third or fourth aspect of the present invention and the input / output signals of the neural network are used as inputs to set parameters such as a coupling coefficient and a threshold of the neural network. It is constituted by a parameter updating means by a genetic algorithm calculating means for calculating and outputting to this neural network.

According to a sixth aspect of the present invention, there is provided a signal processing device, wherein the neural network according to the first, second, third or fourth aspect is used, and the input / output signals of the neural network are used as inputs to set parameters such as a coupling coefficient and a threshold value of the neural network. It is constituted by parameter updating means by random search calculation means for calculating and outputting to this neural network.

The signal processing device according to claim 7 is the signal processing device according to claim 5 or 6 , wherein the signal processing device has a plurality of neural networks.

An autonomous system according to claim 8 is a neural network according to any one of claims 1, 2, 3 or 4 in which a state is dynamically changed, signal input means for inputting a signal to the neural network, and the neural network. And a signal output means for inputting a signal output from.

An autonomous robot according to a ninth aspect generates a signal to be input to the neural network according to the first, second, third or fourth aspect in which the state dynamically transits and the neural network. It is provided with a sensor and drive display means for receiving or outputting a signal output from the neural network as a control signal to drive or display.

According to a tenth aspect of the present invention, there is provided a moving system, which comprises a potential detecting means for detecting a potential of a work area.
A neural network according to any one of claims 1, 2, 3 or 4 which receives a detection output from the potential detecting means, a signal converting means for converting and outputting an output signal from the neural network, and a signal converting means from the signal converting means. It is provided with steering wheels and driving wheels to which output signals are given as control signals.

A moving system according to claim 11 is a potential detecting means for detecting the potential of a work area.
7. A signal processing device according to claim 5 , which receives a detection output by the potential detecting means, and a signal converting means for converting and outputting an output signal from the signal processing device.
It is provided with steering wheels and drive wheels to which output signals from the signal converting means are given as control signals.

A mobile system according to claim 12 is the following:
The parameter updating means in the configuration of the mobile system described in 10 or 11 is provided outside the mobile body and is connected by the communication means.

[0019]

In the neural network according to the first aspect, the displacement amount of the internal activation value is changed according to the neuron output value.
(Sum of products of input value of external input signal and coupling coefficient value) +
(Threshold)-(product of internal activation value and neuron output value)
Since multiple neuron elements with multiple inputs and one output, which are calculated and determined, are connected to each other, it is possible to show dynamic states such as steady state, chaos, limit cycle, etc. Can be easily constructed.

In the neural network according to the present invention, since the internal activation value storage means is provided, a plurality of signals can be input to the same signal line in a time series,
Expandability of the neural network according to claim 1,
The processing speed can be increased and the system flexibility is increased.

In the neural network according to the third aspect, the storage means for storing the threshold value for determining the displacement amount of the internal activation value is also used as the storage means for the coupling coefficient. The scale of neuron elements in the network can be reduced, and the scale of the neural network can be increased.

In the neural network according to the fourth aspect, since the latch means is provided on the output line from the external input means to the adder means, pipeline processing becomes possible, processing speed is improved, and system flexibility is increased. Becomes

In the signal processing device according to claim 5 ,
The input / output signal or the input signal of the neural network is input to the neural network according to claim 1, 2, 3 or 4 as described above, and parameters such as a coupling coefficient and a threshold value of the neural network are calculated to the neural network. Since the parameter updating means by the outputting genetic algorithm computing means is combined, it becomes possible to update the parameters of the neural network,
The applicability of the system is enhanced.

Also in the signal processing device according to claim 6 ,
The input / output signal or the input signal of the neural network is input to the neural network according to claim 1, 2, 3 or 4 as described above, and parameters such as a coupling coefficient and a threshold value of the neural network are calculated to the neural network. Since the parameter updating means by the output random search computing means is combined, the parameters of the neural network can be updated, the applicability of the system is enhanced, and the parameter updating means is simpler than that by the genetic algorithm computing means. Hardware can be easily implemented.

In the signal processing device according to claim 7 ,
Since it has a plurality of neural networks, the processing speed of the signal processing device according to the above-mentioned claim 5 or 6 is improved, and the performance is improved by allowing the input / output signals of a plurality of neural networks to interact with each other. It is also possible.

In the autonomous system according to claim 8 ,
Since the autonomous system is constructed by using the neural network whose state dynamically changes, it is possible to construct a highly autonomous system.

In the autonomous robot according to the ninth aspect, since the autonomous robot is configured by using the neural network whose state dynamically changes, it is possible to provide a highly autonomous robot.

In the mobile system according to the tenth to twelfth aspects, the mobile system is constructed by using a neural network whose state dynamically changes or a signal processing device equipped with such a neural network.
With a very simple structure, it is possible to build a complicated mobile system.

[0029]

EXAMPLES be described with reference to an embodiment of the invention of claim 1, wherein in FIGS. FIG. 1 and FIG. 2 show a configuration example of one neuron element 11 which is an element constituting the neural network of this embodiment showing dynamic states such as periodic output and chaotic output due to internal or external factors. Show.

First, FIG. 2 shows the input / output relationship of one neuron element 11, and parallel input signals x ₁ (t) to x _N (t) (N is Any positive integer, t means time) and the output signal y
It has a multi-input / single-output configuration for outputting (t).

FIG. 1 is a block diagram showing a configuration for internal processing of such a neuron element 11. First, an adder (adding means) 12 for adding a plurality of signals and a non-linear converter (non-linear converting means) 13 for non-linearly converting and outputting the addition result are sequentially provided. To the input side of the adder 12, a plurality of multipliers 14 serving as external input means, a computing unit 15 serving as internal input means, and a memory 16 storing a threshold value θ are connected. The multiplier 14 calculates the product of each input signal x _i (t) (i = 0 to N) and the coupling coefficient w _i stored in the memory 17. Further, the arithmetic unit 15 has an internal activation value u (t) which is an output value of the adder 12.
And the output signal y of the non-linear converter 3 that is the neuron output value
Predetermined arithmetic processing is performed on (t) and. Therefore, in such a neuron element 11, the amount of displacement of the internal activation value u (t) is determined according to the output value of the own neuron element 11.

A concrete calculation method in such a neuron element 11 is as shown in, for example, equations (1) to (3). However, in the case of a discrete system, equation (1) becomes equation (4).

[0033]

[Equation 1]

In these cases, the processing in the computing unit 15 is {1-y (t)} u (t), and the processing in the non-linear computing unit 3 is the equation (2). This corresponds to the arithmetic processing of the invention described in claim 1 .

A plurality of such neuron elements 11 are prepared and connected to each other to form a neural network. For example, as shown in FIG. 3, three neuron elements 11A to 11C are prepared, the inputs and outputs are mutually connected, and, for example, an input signal from the outside is given to the neuron element 11A, and the neuron elements 11B and 11C are supplied. The outputs 1 and 2 are respectively taken out. With such an extremely simple system configuration, a chaotic behavior is shown when the input value is a certain value, and a periodic behavior is shown when the input value is a value, thus constructing a powerful nonlinear system. It is suggested that it can be done. Therefore, such a neural network is a neural network capable of showing a steady state and dynamic states such as chaos and limit cycle. That is, the output value of each neuron element 11 exhibits dynamic behavior such that the output values of the neuron elements 11 dynamically transition to each other due to an external input signal, and a neural network having adaptability to a dynamic state can be constructed. Become.

The non-linear conversion unit 3 may convert the output signal from the adder 12 by using a table lookup, or may calculate the non-linear function by a DSP or a CPU. Also, adder 1
2, the multiplier 14 and the arithmetic unit 15 may also be realized by a DSP or a CPU, and the memory 16 may be a register that holds only one data. Further, such a neuron element 11 may be partially or entirely realized by software.

An embodiment of the invention described in claim 2 will be described with reference to FIG. The same parts as those shown in the above-mentioned embodiments are designated by the same reference numerals (the same applies to the following embodiments). In the above-described embodiment, the neuron element 11 to which the input signal x _i (t) is input in parallel is configured, but in the present embodiment, the input signal x _i (t) is sequentially, that is, in time series. It is configured as an input neuron element 18. Figure 1
In contrast, a memory 19 connected between the input and output of the adder 12 and storing the internal activation value u (t) by the output signal of the adder 12 is added as a storage means. In addition, the calculation result of the calculator 15 and the threshold value θ stored in the memory 16
An adder 10 for adding and outputting to the memory 19 is also added. Further, the input signal x to the adder 12
_{The number} of signal lines of _i (t) is one, and the memory 17 is supposed to hold all the coupling coefficients w _i corresponding to each input, and the input signal x _{i is} controlled by an address controller (not shown). It is configured to output the coupling coefficient w _i corresponding to (t) to the multiplier 14. Or this memory 1
7 may be configured by a shift register.

In such a configuration, the input signal x
_i (t) is sequentially input, and the product of the corresponding coupling coefficient w _i held in the memory 14 is calculated by the multiplier 14. The output value of the multiplier 14 is added by the adder 12 and the sum of the input signals x ₁ (t) to x _i-1 (t) held in the memory 19, and nonlinearly converted by the nonlinear converter 3. After that, the output signal y (t) is output. The final sum u (t-1) of the adder 12 at the previous time (t-1) (the sum after all the input signals x _i (t-1) have been input) and the previous time (t- The output signal y (t-1) of 1) is arithmetically processed by the arithmetic unit 15, is added by the threshold value θ held in the memory 16 and the adder 20, and the obtained sum is loaded into the memory 19 as an initial value. It

As in the case of the neuron element 11, a plurality of such neuron elements 18 are prepared and connected to each other according to FIG. 3 to form a neural network. According to the neural network of the present embodiment, it is possible to input a plurality of input signals x _i (t) on the same line in time series. Since the processing speed can be improved, the flexibility of the system is increased.

An embodiment of the invention described in claim 3 will be described with reference to FIG. In the present embodiment, in the configuration of the embodiment shown in FIG. 4, the memory 17 that stores and holds the coupling coefficient w _i also serves as a storage unit that also stores the threshold value θ for determining the internal displacement amount u (t). Is configured. That is,
The threshold θ is also stored in the memory 17, and the threshold θ is stored in the memory 17.
If the multiplier 14 is such that the input signal becomes 1 when the threshold value θ is output, the threshold value θ is directly input from the multiplier 14 to the adder 12. Become. Therefore, the neuron element 21 of the present embodiment
Is the memory 16 from the neuron elements 18 shown in FIG.
And the adder 20 is omitted.

As in the case of the neuron element 11, a plurality of such neuron elements 21 are prepared and connected to each other according to FIG. 3 to form a neural network. According to the neural network of the present embodiment, in addition to the effects of the above-described embodiment, the scale of the neuron element 21 can be reduced, so that the neural network can be constructed in a large scale.

An embodiment of the invention described in claim 4 will be described with reference to FIG. In this embodiment, a latch 22 is added to the output line of the multiplier 14 in the configuration of the neuron element 18 shown in FIG. 4 to form a neuron element 23. The presence of the latch 22 enables pipeline processing, and the multiplication processing by the multiplier 14 and the addition processing by the adder 12 can be performed at the same time.

As in the case of the neuron element 11, a plurality of such neuron elements 23 are prepared and connected to each other according to FIG. 3 to form a neural network. According to the neural network of the present embodiment, in addition to the effect of the above-described embodiment using the neuron element 18, the processing speed of the neuron element 21 can be improved and the flexibility of the neural network system can be increased. it can.

As shown in FIG. 7, a latch 22 may be provided on the output line of the multiplier 14 in the neuron element 21 shown in FIG. 5 to form the neuron element 24 (claim 3). It corresponds to the invention described in claim 4 with respect to the described invention).

As described above, the neuron elements 11 and 1 having various configurations
Although the embodiments of the neural network using 8, 21, 23, and 24 have been described, these neural networks can output various signals depending on parameters such as the coupling coefficient w _i and the threshold value θ. Therefore, in applying such a neural network, an embodiment relating to a signal processing device for determining parameters such as a coupling coefficient w _i and a threshold value θ for outputting a signal suitable for the application will be shown below.

First, an embodiment of the invention described in claim 5 will be described with reference to FIG. In this embodiment, a genetic algorithm (Genetic Algorithm) is used for updating the parameters. The signal processing device 25 of this embodiment is a neural network (NN) according to any of the embodiments described above.
W) 26 (that is, any one of the neuron elements 11, 18, 21, 23 or 24 is used as a neuron element) and a genetic algorithm computing means (GA) 27 which serves as a parameter updating means. There is. The neural network 26 receives a signal from the outside world through an input line 28 and outputs the signal to the outside world through an output line 29. The input side of the neural network 26 is connected to the input line 28 and the output line 29 of the genetic algorithm calculating means 27, and the output line 30 of the genetic algorithm calculating means 27 is connected to the neural network 26.
It is connected to the.

With such a configuration, the neural network 26 inputs a signal from the outside through the input line 28,
The processed signal is output from the output line 29 to the outside world. At this time, the genetic algorithm computing means 27 inputs signals from these input lines 28 and output lines 29 and calculates and updates parameters such as the coupling coefficient w _i and the threshold value θ from these signals,
The updated parameters are loaded from output line 30 into neural network 26.

At this time, the genetic algorithm computing means 27
The genetic algorithm calculation process performed by will be described. 1. Coding a coupling coefficient w _i and a threshold θ into a gene. That is, these values are used as the loci values. Generate 2 M genes. Genes are sequentially loaded into the 3 neural network 26, and an input line 28 and an output line 29 when each gene is used
Calculate the fitness from the above signal. 4 M genes are selected and stored with a probability according to the fitness. However, the L highest fitness values are preferentially stored. Cross 5 genes. Mutate a gene with a probability of 6 p. The locus causing mutation is added to the original value by the stochastic mutation amount δ (δ is a real number). Return to 7 3 . If the end conditions are met, the process ends.

According to this embodiment, the neural network 26 is provided with the genetic algorithm calculating means 27.
Since the parameters used by can be updated, the applicability as a neural network system can be enhanced.

An embodiment of the invention described in claim 6 will be described with reference to FIG . In this embodiment, random search is used to update the parameters. Signal processing device 3 of this embodiment
Reference numeral 2 is composed of a neural network (NNW) 16 according to any of the above-described embodiments and a random search calculation means 33 serving as a parameter updating means. That is, the random search calculation means 33 is provided in place of the genetic algorithm calculation means 27 in FIG. 8, and the input side 28 and the output line 29 of the neural network 26 are connected to the input side of the random search calculation means 33. The output line 34 of the random search calculation means 33 is connected to the neural network 26.

With such a configuration, the neural network 26 inputs a signal from the outside world through the input line 28,
The processed signal is output from the output line 29 to the outside world. At this time, the random search calculation means 33 receives these input lines 28,
The signals from the output line 29 are input, parameters such as the coupling coefficient w _i and the threshold value θ are calculated and updated from these signals, and the updated parameters are loaded from the output line 34 to the neural network 26.

At this time, the random search calculation process performed by the random search calculation means 33 will be described. One coupling coefficient w _i and threshold θ are set as one set. Generate 2 M sets. The sets are sequentially loaded into the 3 neural network 26, and the fitness is calculated from the signals on the input line 28 and the output line 29 when each set is used. 4 M sets are selected and stored with a probability according to the fitness.
However, the L highest fitness values are preferentially stored. Mutate the set with a probability of 5 p. The elements that cause mutation are added to the original value by the stochastic variation amount δ (δ is a real number). Back to 6 3. If the end conditions are met, the process ends.

According to the present embodiment, the random search operation means 33 is provided and the parameters used by the neural network 26 can be updated, so that the applicability as a neural network system can be enhanced. In particular, since the parameter updating means can be simplified as compared with the case of using the genetic algorithm computing means 27, the hardware implementation becomes easier.

Further, these signal processing devices 25, 31,
Both 32 and 35 have one neural network 2
Although there is a configuration using a 6 may be configured to use in parallel with the genetic algorithm computing means 27 or random search operation means 33 these neural network by providing a plurality of neural networks ( It corresponds to the invention of claim 7 ). According to this, the optimum value of the parameter can be searched at high speed, and the processing speed can be improved. Especially when there are interactions between the input and output signals of multiple neural networks,
It can be expected to improve performance.

Next, starting with the above-mentioned neural network 16, the point is to show an example of use of a neural network whose state dynamically changes in the following embodiments. First, an embodiment of the invention described in claim 8 will be described. In this embodiment, a neural network whose state dynamically changes, for example, the above-mentioned neural network 26 is used to construct an autonomous system with high autonomy. That is, the neural network 26
The system configuration has a signal input means for inputting a signal to the input terminal and a signal output means for inputting a signal output from the neural network 26, and an output which dynamically transits to a certain input is generated.

Further, a reference configuration example of the autonomous system is shown in FIG.
13 to FIG. In this configuration example, a chaotic neural network as shown below is used instead of the neural network 26 in constructing. FIG. 10 shows a configuration example of one neuron element 40 which is an element constituting the chaotic neural network of this configuration example. First, an adder (adding means) 41 for adding a plurality of signals, a non-linear conversion section (non-linear conversion means) 42 for non-linearly converting the addition result, and an output section (output means) 43 for outputting the conversion result to the outside are provided. They are provided in order.
A plurality of external input sections (external input means) 44, a first internal input section (first internal input means) 45, and a second internal input section (second internal input means) are provided on the input side of the adder 41. ) 46
And a memory (storing means) 47 storing a predetermined input signal
Are connected in parallel, and each signal can be input to the adder 41. As a chaotic neural network, a plurality of such neuron elements 4
It is configured by connecting 0s to each other.

Here, the external input section 44 is, for example, as shown in FIG.
As shown in FIG. 11 , the multiplier 49 multiplies the external input signal and the signal stored in the memory 48, specifically, the coupling coefficient, and inputs the multiplication result to the adder 41. The first internal input unit 45 is, for example, a diagram
12 , the output signal 50 from the adder 41 and a signal stored in the memory 51, specifically, a predetermined coefficient are multiplied by a multiplier 52, and the multiplication result is sent to the adder 41. It is something to input. The second internal input unit 46 includes a nonlinear conversion unit 4 as shown in FIG.
The output signal 53 from 2 and a signal stored in the memory 54, specifically, a predetermined coefficient are multiplied by a multiplier 55, and the multiplication result is input to the adder 41. The output unit 43 is composed of a latch that holds the output value as the conversion result from the nonlinear conversion unit 42 for a certain period.

The non-linear conversion section 42 may convert the output signal from the adder 41 by using a table lookup, or may calculate the non-linear function by a DSP or a CPU. The above-mentioned adder 41 and multipliers 49, 52, 55 are also DSPs or CPs.
It may be realized by U, and the memory 4
7, 48, 51 and 54 may be registers that hold only one data. In addition, such a neuron element 40
May be realized partially or entirely by software.

According to such a chaotic neural network configuration, the output value of each neuron element 40 dynamically transits depending on the value stored in the memories 47, 48, 51 and 54 and the value of the external input signal. It exhibits a dynamic behavior, and a neural network having adaptability to a dynamic state can be constructed. Therefore, even when such a chaotic neural network is used, an autonomous system with high autonomy can be constructed as in the case of using the neural network 26. That is, a system configuration having a signal input means for inputting a signal to a chaotic neural network and a signal output means for inputting a signal output from the chaotic neural network is provided, and an output dynamically transiting to a certain input is generated. It is a thing.

Here, each neuron element 4 shown in FIG.
In the structure of 0, but it may also be to remove the memory 47. In addition, the first embodiment in which the memory 48 and the multiplier 49 are substantially eliminated
Output signal 5 from adder 41 with internal input section 44 configuration
0 but it may also be so as to directly input directly to the adder 41. Similarly, substantially eliminating the second internal input section 45 constituting the memory 51 and the multiplier 52, but it may also be so as to directly input to the direct adder 41 output signal 53 from the nonlinear conversion unit 42. According to these, the neural network can be constructed more easily.

Next, another reference configuration example is shown in FIG. 14 and FIG.
This will be described with reference to 15 . This configuration example uses a genetic algorithm to learn a chaotic neural network 56 having the neuron element 40 described in the above configuration example as a component, that is, the memories 47, 48, 51 in each neuron element 40. The value of 54 is updated. First, the method of applying the genetic algorithm will be described. As shown in FIG. 14 , a gene whose value is the value of the memory of each neuron element 40 is assumed. Here, the memory 48 corresponds to the coupling coefficient, and there are as many external input units 44 as the neuron element 40.

In the genetic algorithm, necessary data are
As described above, it is coded by considering it as one gene (individual). A plurality of such individuals are prepared to form a group. Then, the processing is performed according to the following procedures 1 to 7 . Number to generate 1 gene (individuals). 2 Each individual is loaded into the chaotic neural network 56, and the evaluation value Er is calculated from the output value for the training data. Here, the closer the evaluation value Er is to 0, the better the evaluation. 3 Use evaluation values, and "select" if they are above a certain standard value. 4 Perform "crossover" between the remaining individuals. 5 Create new individuals by the number of selected individuals. 6 "mutate" individuals at a certain rate. 7 The above 2 to 6 are repeated, and if the optimum individual is found, the process ends. The evaluation criteria and the mutation rate are arbitrarily determined empirically.

FIG . 15 shows a device configuration for performing learning using such a GA, in which an evaluator (evaluation means) 57 is connected between the input and output of the chaotic neural network 56 and the chaotic neural network. GA calculator for 56 (genetic algorithm calculator)
58 is connected and comprised.

In such a configuration, the evaluator 57 inputs the test data as an input signal to the chaotic neural network 56 and inputs the output result obtained from the chaotic neural network 56. Therefore, the evaluator 57 uses the evaluation result Er from the output result of the chaotic neural network 56 and the given input signal to evaluate the value Er.
To calculate. The GA computing unit 58 writes and updates the values of the memories 47, 48, 51, 54 in the chaotic neural network 56 based on the calculated evaluation value Er.

According to this configuration example, since the learning of the chaotic neural network 56 is possible, the flexibility becomes higher. Therefore, the flexibility of the autonomous system configured by using the chaotic neural network 56 is increased.

Next, another reference configuration example will be described. This configuration example also relates to the learning method of the chaotic neural network 56 having the configuration described in the above configuration example. However, the local learning rule such as Hebb's learning rule causes the memory 48 in the external input unit 44 to store data in the memory 48. The value (coupling coefficient value) is updated. That is, after updating the memories 48, 51, 54, the values before updating are MEM _NEW , MEM _OLD , and each memory value M.
Letting the input value multiplied by EM _OLD be IN, MEM
The value of the memory 48 is updated according to the equation _NEW = MEM _OLD + A * (IN-B).
Here, the values of A and B are empirically determined arbitrarily for each memory 48, 51, 54 in the chaotic neural network 56. In the learning method of this configuration example, the memory 4
The value of 7 is a fixed value.

Also in the case of the present configuration example, since the chaotic neural network 56 can be learned, the flexibility becomes higher. Therefore, the flexibility of the autonomous system configured by using the chaotic neural network 56 is increased.

Next, a reference configuration example will be described with reference to FIG. 16 as a more specific example of this autonomous system. Main structure
Adult example, neural network dynamically state transitions, for example, which is constituted an autonomous robot 59 using a chaotic neural network 56 described with Figures 10 described above with reference to FIG. 15. That is, the signals from the sensors 60 of the autonomous robot 59 are input to the chaotic neural network 56, and the output signals from the individual neuron elements 40 in the output stage are used as control signals for the drive systems or displays of the autonomous robot 59. Here, the display (driving / displaying means) 61 is provided to drive or display.

Here, the chaotic neural network 2
6 performs learning and self-organization in accordance with the above-described learning rule and evolves, and further, because the state transitions dynamically, it shows different movements even for signals from the same sensor 60. Therefore, a robot having adaptive autonomy can be realized.

In the construction of the autonomous robot 59 of this embodiment, instead of the chaotic neural network 56, the neural network 26 having the neuron element 11 (or 18, 21, 23 or 24) as a constituent element is used. It may be used (corresponding to the invention of claim 9 ).

FIG. 17 shows an embodiment of the invention described in claim 11 .
Through FIG. 20 . In this embodiment, the mobile system is a signal processing device as described above with reference to FIGS.
This is a configuration of an autonomous traveling robot 62 using 25 or 32 . First, FIG. 17 shows this autonomous mobile robot 6.
2 is a schematic diagram showing the external configuration of the robot main body 63.
Are configured to be movable by front wheels (steering wheels) 64 and rear wheels (driving wheels) 65. Also, the robot body 6
A sensor (potential detection means) 66 such as an electrometer and an altimeter for sensing the potential (potential, altitude, etc.) of the work area is mounted on the unit 3.

FIG . 18 shows such an autonomous mobile robot 62.
8 is a block diagram configuration of a controller system 67 including a neural network 26 (the neural network 26 itself has the configuration illustrated in FIG. 3) and a parameter updating unit 27 or 33 as shown in FIGS. 8 and 9 . (Substantially equivalent to the signal processing device 25 or 32 ) is provided and is configured to capture the potential signal from the sensor 66. A decoder (signal conversion means) 68 is provided on the output side of the controller system 67 and is divided into two signals for the front wheels 64 and the rear wheels 65. The front wheel 64 is a signal indicating the amount of movement in the left-right direction, the rear wheel 65 is a signal indicating the amount of movement in the front-rear direction, and the decoder 68 converts the output from the controller system 67 into a steering angle and a drive amount. Front wheel 64, rear wheel 6
5 is actually controlled.

The autonomous mobile robot 62 having such a configuration
Let us consider a case where is quickly moved to a region having the highest potential in the work region. Therefore, the sum of potential inputs in the work area is used as the fitness used for updating the parameters. FIG. 19 shows an example of the potential of the work area.

If the autonomous mobile robot 62 is left in such a work area, the system automatically moves to a place having the highest potential. A movement locus T of the autonomous mobile robot 62 as described above is, for example, as shown in FIG. 20, and a chaotic search (random search) is performed on the potential flat land A, and the hill climbing method is performed from the foot C of the potential mountain B. The target is searched (the steepest descent method), reaches the top D of the potential, and stays in the vicinity. That is, the movement trajectory T shown in FIG. 20 is a simulation for controlling the autonomous traveling robot 62 by the neural network 26 formed by interconnecting the three neuron elements 11A to 11C as shown in FIG. 3, for example. This is due to the results of the generational robots. A genetic algorithm was used to update the parameters such as the coupling coefficient.

That is, in a moving system having only a single sensor such as the sensor 66, it is very difficult with ordinary software to execute a task of moving to a region having the highest automatic potential, but this is difficult. According to the embodiment, for example, only three neuron elements 11A
A non-equilibrium neural network of ~ 11C can realize a highly non-linear task.

In this configuration example, the autonomous traveling robot 62 using the signal processing device 25 or 32 provided with the parameter updating means 27 or 33 has been described, but it is configured using only the neural network 26. It may be (corresponding to the invention of claim 10 ).

FIG. 21 shows an embodiment of the invention described in claim 12 .
Will be described. In the above embodiment, the autonomous traveling robot 6
Although the controller system 67 mounted in 2 is configured as an offline system including the parameter updating means 27 or 33, in the present embodiment, the controller system 69 using only the neural network 26 is used, and the parameter updating means 27 or 33 is a remote host. The computer has an online system in which signals are transmitted and received by the communication means 70. In such an online system, the individual used by the parameter updating means 27 or 33 may be assigned to a plurality of robots (autonomous traveling robot 62).

[0078]

According to the neural network of the first aspect of the invention, the displacement amount of the internal activation value (the product of the input value of the external input signal and the coupling coefficient value) is changed according to the neuron output value.
Sum of) + (threshold value)-(internal activity value and neuron output value)
Since a plurality of multi-input / one-output neuron elements, which are calculated and determined as (the product of), are connected to each other, it is possible to show a steady state and dynamic states such as chaos and limit cycle. A complex system can be easily constructed in a network.

According to the neural network of the invention described in claim 2, since it has a storage means for the internal activation value, it is possible to input a plurality of signals in time series to the same signal line. The expandability and processing speed of the neural network of the invention can be improved, and the flexibility of the system can be improved.

According to the neural network of the third aspect of the present invention, the storage means for storing the threshold value for determining the displacement amount of the internal activation value is also used as the storage means for the coupling coefficient. The size of the neuron element in the neural network of the described invention can be reduced,
It is possible to promote the scale-up of the neural network.

According to the neural network of the invention described in claim 4, since the latch means is provided on the output line from the external input means to the adder means, the pipeline processing becomes possible, the processing speed is improved, and the system flexibility. The property can be further improved.

According to the signal processing apparatus of the fifth aspect of the invention, the input / output signal or the input signal of the neural network is input to the neural network of the first, second, third or fourth aspect of the invention as described above. As the parameter updating means by the genetic algorithm computing means for calculating the parameters such as the coupling coefficient and the threshold value of the neural network and outputting them to the neural network is combined, the parameter updating of the neural network becomes possible. Applicability can be increased.

Also in the signal processing apparatus of the invention described in claim 6 , the input / output signal or the input signal of this neural network is input to the neural network of the invention described in claim 1, 2, 3 or 4 as described above. Since the parameters of the neural network such as the coupling coefficient and the threshold value are calculated and output to the neural network, the parameter updating means by the random search operation means is combined, so that the parameters of the neural network can be updated. In addition, the parameter updating means is simpler than the case of using the genetic algorithm computing means, and the hardware implementation can be facilitated.

According to the signal processing device of the invention described in claim 7, since it has a plurality of neural networks, the processing speed of the signal processing device of the invention described in claim 5 or 6 can be improved. It is possible to improve the performance by allowing the input / output signals of a plurality of neural networks to interact with each other.

According to the autonomous system of the eighth aspect of the present invention, since the autonomous system is constructed using the neural network whose state dynamically changes, it is possible to construct a highly autonomous system.

According to the autonomous robot of the ninth aspect of the invention, since the autonomous robot is configured by using the neural network whose state dynamically changes, it is possible to provide a highly autonomous robot.

According to the mobile system of the tenth to twelfth aspects of the present invention, the mobile system is configured by using a neural network whose state dynamically changes or a signal processing device equipped with such a neural network.
With a very simple structure, it is possible to build a complicated mobile system.

[Brief description of drawings]

1 is a block diagram of one neuron element showing one embodiment of the invention of claim 1, wherein.

FIG. 2 is a schematic diagram showing an input / output relationship of one neuron element.

FIG. 3 is a schematic diagram showing a configuration example of a neural network.

FIG. 4 is a block diagram of one neuron element showing an embodiment of the invention described in claim 2;

FIG. 5 is a block diagram of one neuron element showing an embodiment of the invention according to claim 3;

FIG. 6 is a block diagram of one neuron element showing an embodiment of the invention described in claim 4;

FIG. 7 is a block diagram of one neuron element showing a modification thereof.

FIG. 8 is a block diagram of a signal processing device showing an embodiment of the invention described in claim 5 ;

FIG. 9 is a block diagram of a signal processing device showing an embodiment of the invention described in claim 6 ;

FIG. 10 is a block diagram of one neuron element showing a reference configuration example .

FIG. 11 is a block diagram showing the configuration of the external input unit .

FIG. 12 is a block diagram showing a configuration of a first internal input section thereof .

FIG. 13 is a block diagram showing a configuration of a second internal input section thereof .

FIG. 14 is an explanatory diagram of a gene showing another reference configuration example .

FIG. 15 is a block diagram.

FIG. 16 is a schematic view showing an embodiment of the invention according to claim 9 ;
It is a block diagram.

FIG. 17 is an autonomous view showing an embodiment of the invention according to claim 11;
It is a schematic block diagram of a traveling robot .

FIG. 18 is a block diagram thereof .

FIG. 19 is a characteristic diagram showing an example of a potential of a work area .

FIG. 20 is a plan view showing a movement trajectory in a work area .

FIG. 21 is a block diagram showing an embodiment of the invention according to claim 12;
Tsu is a click view.

FIG. 22 is a block diagram of a chaotic neuron showing a conventional example .

[Explanation of symbols]

11 neuron element 12 adding means 13 non-linear converting means 14 external input means 15 internal input means 17 storage means 18 neuron element 19 storage means 21 neuron element 22 latches 23, 24 neuron element 25 signal processing device 26 neural network 27 genetic algorithm computing means parameter updating unit 32 signal processing device 33 random search operation means by the parameter update manual stage 5 9 autonomous robot 60 sensor 61 drive display unit 62 moves the system 63 moves body 64 steering wheel 65 the drive wheel 66 potential detection means 67 signal processing device 68 according to the Signal conversion means 69 Signal processing device 70 Communication means

Continuation of the front page (56) References Tokumeihei 4-503876 (JP, A) Shoichi Sawada et al., “On Neural Networks Based on Nagumo-Sato Cell”, IEICE Technical Report, Japan, Japan Society The Institute of Electronics, Information and Communication Engineers, March 18, 1993, Vol. 92, No. 521 (NC92-95 to 127), pp. 151-158 Shunichi Amari (Translated), "PDP Model", Japan, Sangyo Tosho Co., Ltd., February 27, 1989, first edition, pp. 325-334, ISBN: 4-7828-5125-1 Kazuo Asakawa, "Learning Control of Robots by Neurocomputers", Computer Control, Japan, Corona Co., Ltd., October 10, 1988, No. 24, pp. 108-115, ISBN: 4-339-02043-5, S. Shimizu, et al., "Electronic Circuit Model of Chaotic Neural Network," IEICE Transactions, Japan, The Institute of Electronics, Information and Communication Engineers, March 1990. 25th, Vol. J73-A, No. 3, pp. 495-508, JST Material No .: S0621A Toshimitsu Ushio, "Chaos synchronization in one-dimensional one-way coupling chaos synchronization, Technical Report of IEICE, Japan, The Institute of Electronics, Information and Communication Engineers, October 1993. 20th, Vol.93, No.278, pp.21-26, JST Material No .: S0532B (58) Fields investigated (Int.Cl. ⁷ , DB name) G06N 1/00-7/08 G06G 7/60 JST file (JOIS) CSDB (Japan Patent Office)

Claims (12)

(57) [Claims] 1. An addition means for adding a plurality of signals, a non-linear conversion means for non-linearly converting an internal activation value resulting from the addition of the addition means, and outputting the conversion result to the outside as a neuron output value, and an external input signal. And an external input means for multiplying the coupling coefficient and outputting the multiplication result to the adding means, and an internal input means for performing arithmetic processing on the internal activation value and the neuron output value and outputting to the adding means. displacement of said internal activity value in response to the neuron output value (outside
Partial input signal input value and coupling coefficient value product sum) + (threshold
Value)-(product of internal activation value and neuron output value)
A neural network, characterized in that a plurality of neuron elements having multiple inputs and one output, which are calculated and determined, are connected to each other. 2. The neural network according to claim 1, further comprising storage means connected to the addition means for storing the internal activation value. 3. The neural network according to claim 2, wherein the storage means for storing the threshold value for determining the displacement amount of the internal activation value is also used as the storage means for the coupling coefficient. 4. A latch means is provided on the output line from the external input means to the adder means.
Neural network described. 5. The neural network according to claim 1, 2, 3 or 4 , and the input / output signals of this neural network as input, and calculates parameters such as coupling coefficient and threshold value of said neural network and outputs to this neural network. A signal processing device comprising: a parameter updating unit by a genetic algorithm computing unit. 6. The neural network according to claim 1, 2, 3 or 4 , and the input / output signals of this neural network as input, and calculates parameters such as coupling coefficient and threshold value of said neural network and outputs to this neural network. A signal processing device comprising a parameter updating means by random search computing means. 7. The signal processing device according to claim 5 , further comprising a plurality of neural networks. 8. A dynamic state transition, 1, 2, 3
Alternatively , an autonomous system comprising: the neural network according to item 4; signal input means for inputting a signal to the neural network; and signal output means for inputting a signal output from the neural network. 9. The state transition according to claim 1, 2, 3
Or a neural network according to 4, and a sensor that generates a signal to be input to this neural network,
An autonomous robot comprising: a drive display unit that receives or outputs a signal output from the neural network as a control signal. 10. A potential detecting means for detecting a potential of a work area, a neural network according to claim 1, 2, 3 or 4 which receives a detection output by this potential detecting means, and an output signal from this neural network. A signal converting means for converting and outputting
A moving system having steering wheels and driving wheels to which output signals from the signal converting means are given as control signals. 11. A potential detecting means for detecting a potential of a work area, a signal processing device according to claim 5 or 6 which receives a detection output by the potential detecting means, and an output signal from the signal processing device. And a steering wheel and driving wheels to which output signals from the signal converting means are given as control signals. 12. The mobile system of claim 10 or 11, wherein the parameter update means is connected by communication means provided outside the mobile body.