RU2475843C1 - Adaptive control device, neuron-like base element and method for operation of said device - Google Patents

Abstract

FIELD: information technology.

SUBSTANCE: device includes a plurality of neuron-like base elements (BE) which are grouped into modular units such that outputs of like BE are orthogonal to inputs of other BE and are separated by a dielectric film with a semiconductor layer. The neuron-like BE has a multipoint input for reception and summation of electrical signals, a working output and a signal processing unit, having a threshold voltage former, a comparator and two normalised voltage formers.

EFFECT: simple design of the device, broader functional capabilities thereof and a more complex category of problems solved.

10 cl, 7 dwg

Description

The invention relates to analog-digital control devices and can be used to create complex multi-parameter automatic control systems for various objects and technological processes that allow the object to change its reaction depending on changes in the nature of external influencing factors, in pattern recognition systems, in robotics, and also for human brain modeling.

According to the general theory of systems, the control device systematically or as necessary develops control actions on a certain process. Self-organizing control devices are based on the principles of optimizing the solution of tasks, for which the input information is processed in accordance with a set of rules in order to obtain control signals that adequately solve the task.

The best example of self-organization is the human brain. To implement devices that function by analogy with the operation of neural networks, numerous technical solutions have been created, mainly in the field of computer design.

The work of adaptive systems is organized using an artificial neural network, which to one degree or another models the nervous system of living organisms and operates on the basis of neural-like elements. Devices for modeling neural networks can be analog-to-digital, hardware or computer. The first group includes devices in which models of individual neurons and interneuronal connections are realized in the form of electronic elements using a multi-element analog or digital base. Common disadvantages of such devices are their hardware complexity and high cost, as a result of which it is impossible to create models of large neural networks with the qualities of multifunctionality, difficulties in forming and building up complex neural structures. The second group includes devices in which models of neurons or models of interneuronal connections are made by various devices of functional electronics, for example, optoelectronic models. Common disadvantages of these models are lack of flexibility, the complexity of restructuring the network structure, the complexity of changing the functional characteristics of neurons and interneuronal connections. Another disadvantage inherent in most such devices is the unsatisfactory state of the production technology of the element base. The third group includes universal and specialized programmable computing devices that are used to model neural networks. At the same time, flexibility is achieved in changing interneuronal connections and characteristics of neurons; large neural networks can be modeled. The main disadvantages of such devices are the bulkiness of the software and low performance.

A large number of neural networks of various types are known (see Artificial Neural Networks: Concepts and Theory, IEEE Computer Society Press, 1992; J. Anderson, An introduction to neural networks, Chapter 2, MIT, 1999), whose common drawback is the difficulty in determining weights synaptic connections and their correction, the method of specifying which either roughly simplifies the real processes or complicates the design so much that its actual implementation is hardly possible.

A neural-like element is known on the basis of which it is possible to construct a device for solving a class of problems of evaluating the functioning of open complex systems, assessing the degree of optimality of various solutions (RU 2295769, publ. 2007). The system has limited functionality, in particular, is incapable of self-learning.

A number of technical devices are known for controlling complex technological processes (Neural network, RU 66831, 2007; Domain neural network, RU 7208, 2008; Modular computing system, RU 75247, 2008; Associative memory, RU 77483, 2008, etc.), common disadvantages which are limited functionality and the inability to self-learn, since switching interneuronal connections and transferring information between neurons is based on deterministic connection tables, and for a number of devices it is necessary to use external memory.

Dynamic synapses for signal processing in neural networks are known. The system contains many interconnected basic neurons and is designed to process signals arriving in the form of a sequence of pulses. The essence of the device is the establishment of feedback between the presynaptic and postsynaptic neurons and the use of inhibitory neurons, which continuously changes the signal processing procedure by the synapse, so that when the same signal is presented again, the synapse takes into account the presynaptic picture and gives a different picture at the output. A dynamic change in the conductive structure of a feedback network gives an increase in the computing power of the network at the neural level and an increase in the ability of the network to dynamically learn (US 6643627, 2003).

In the known device, the correction of the information signal is carried out at the local level of pre-synapse - postsynapse, which is effective only for networks of small volume. When the number of layers increases to 4 or more, the useful properties of the dynamic synapse sharply decrease (see Fig. 6a of the patent description). The class of problems solved by such a system is limited; the best results are obtained for the problems of filtering noise in multi-parameter signals and speech analysis, for example, when recognizing the same words spoken by different people with different accents.

There is a known nanotechnological neural network and a method for organizing its work, which in many aspects models the biological brain and the processes of neuron interaction occurring in it (US 7426501. Nanotechnology neural network methods and systems, 2008). The device contains many artificial neurons with nano-processes, suspended and capable of moving in a special environment (dielectric solution). Neighboring adaptive synaptic elements are separated by gaps across which an electric field can be applied, under the influence of which electrical connections are established. Connections are able to amplify, weaken, and even dissolve in a dielectric solution to forget the wrong reactions to input signals, which ensures fast learning of the neural network. For this, the gap between neighboring neurons can be equipped with the appropriate logic.

The known device solves a number of the same problems as the claimed technical solution, but in other ways difficult for physical implementation and practical use.

Known technical nervous system RU 2128857, publ. 1999. The device contains a technical brain, output signal distributor, receptor power regulator, an active learning device, technical brain output receptors, external system receptors, power distribution and training. The neural network is built in the form of a matrix of adjustable resistor elements. At each step of the training, a situation is determined in which the ratio of the deviation of the control signal to the permissible deviation of this signal in the given situation was the largest, and external conductivities of the matrix elements of the technical brain are adjusted for this situation. The system has limited functionality, is difficult to manufacture and contains a large number of external regulators for the correction of reactions, which limits its ability to adapt.

The closest to the claimed technical solution for designation and essential features is a multilayer perceptron device (US 5220641, Multi-layer perceptron circuit device, 1993). The device is selected for the prototype.

The architecture of an adaptive neural network in the form of a multilayer perceptron, similar to the invention of US 5220641, is used most often. Typically, each neuron constructs a weighted sum of its inputs, corrected in the form of a term, and then passes this activation value through the transfer function, and thus the device generates an output signal value. Learning a recurrent neural network essentially consists in examining the surface of errors and finding a global minimum.

The prototype device contains many neural-like basic elements, many connections between which are provided with a special topology, in which the elements are grouped into modular blocks so that the outputs of the BE of one block are located orthogonally or at some angle to the inputs of the BE of another block, which allows you to set between the output of one neural-like element and inputs of many other elements of synaptic communication. Each of the synapses receives signals from one of the signal input lines, calculates the weight of this signal with the prescribed value, and then transfers the result of calculating the weight to one of the signal output lines. Note that since the prototype is the main element in the prototype synapse, hereinafter the input line for the synapse of the prototype is the output line of the neural-like element according to traditional ideas, and the exit line for the synapse of the prototype is the input line for the traditional neuron and for the claimed neuron-like element . Each synapse contains a weighting unit. Self-learning is based on the backpropagation algorithm. The same principles are laid down in a multilayer neural network on MOS inverters according to US 5347613, publ. 1994, in which a new five-stage method of training was announced, but it is based on known methods for correcting bond weights, and training is a modification of the training method according to the invention prototype.

The common disadvantages of self-learning systems by varying, calculating and correcting bond weights are unacceptably long processing times for the input signal configuration to obtain adequate control signals.

According to the prototype, in order to train a neural network to solve a problem, it is necessary to repeatedly correct the weights of each element in such a way that the error decreases - the discrepancy between the actual and the desired output. At the same time, the learning time is growing much faster than the size of the network. The disadvantages of the prototype are the complexity of the device due to the presence of a large number of regulatory elements and its limited functionality.

To eliminate the shortcomings in the multifunctional adaptive control device, it is advisable to abandon the correction of bond weights, including the error back propagation algorithm. In the invention, a different principle was used: instead of infinitely adjusting the weights of the connections, they continuously lay more and more new connections between neurons, “forgetting” the old ones and fixing those that take into account previous events and implement the “correct” reactions to the changing configuration of the input signals. Such an extensive learning path is more promising for adaptive systems and more consistent with the biological nature of intelligence. According to some estimates, the human brain has a potential of hundreds of terabytes, of which about half are involved, and its information content varies on average at a rate of hundreds of kilobytes per second.

According to the invention, the control device can be implemented using a simple node - a neural-like basic element that can form electrical connections with its neighbors and has some basic properties of a biological neuron: the ability to be excited, tired, restore working capacity and energized to transmit a signal without attenuation. To implement such a device, it is necessary to create a multilayer neural network with a very large amount of memory, which is achieved by the special neural network topology described below.

It should be noted here that the Korean company Samsung announced the creation of a substrate for the production of semiconductors, the thickness of which is only 0.08 mm. This is 20% less than the 0.1 mm substrate Samsung had previously released (http://habrahabr.ru/blogs/hardware/14025/). The capacity of the human brain is about 400-500 · 10 ¹² bytes. It can be shown that with already mastered microelectronics technologies in the same volume as the human brain occupies (approximately 1.5 dm ³ = 1500 cm ³ ), at least 5.5 terabytes of information, or about 1% of the capacity, can be placed in the structures of the claimed device brain.

An object of the invention is the creation of an adaptive control device that combines simplicity of design with wide functional capabilities, the creation of a neural-like element (artificial neuron), on the basis of which it is possible to build a trained neural network, and the creation of a method for organizing the work of an adaptive control device with a neural network.

The first of the tasks is solved by the fact that in the known prototype device containing many neural-like basic elements, each of which has an input and an output, it is able to receive and summarize electrical signals from sensors or from other BEs and go into an excited state with the generation of an electrical signal, which can be transferred to many other BEs or to an actuator, grouped in M> 1 flat modular blocks with N> 1 BEs in each so that the BE outputs of one block are located orthogonally or under otorrhea angle to the input BE of another block, forming at least one grid with N nodes, a new design using a multilayer neural network.

To solve the problem, the outputs of the BE of one block are superimposed on the inputs of the BE of another block through a dielectric film with a semiconductor layer, each of these grid nodes, when creating a potential difference sufficient for the breakdown of the dielectric, has the ability to form a resistive connection between the output of one BE and the input of another similar BE whose conductivity decreases with time, the inputs of some arbitrary BEs are connected to the sensors and are assigned to receive signals from external sources, the outputs of some arbitrary BEs are connected s with actuators of the control device and are assigned for the issuance of control signals, and some of the mentioned resistive connections are laid in advance in the manufacture of the control device and when the signal is supplied, it forms at least one electric circuit from the input connected to the sensor to the output connected to the executive mechanism.

Another technical task is the creation of a neural-like basic element suitable for creating self-learning neural networks and the physical implementation of control devices for various purposes.

The technical problem is solved as follows.

Like well-known analogues, a neural-like basic element (BE) of an adaptive control device contains a multi-point input for receiving and summing electric signals from similar basic elements (BE) or from external sources, a working output and a signal processing unit that can generate electrical signals when excited BE. In the basic element according to the invention, the signal processing unit comprises a threshold voltage driver connected to a source of adjustable reference voltage, a comparator and two normalized voltage drivers (VFD), the first of which generates positive rectangular pulses, the second generates negative rectangular pulses of the same size and duration, the comparator is connected to the multi-point input of the BE and with the output of the threshold voltage driver, the input of the first low voltage device is connected to the output of the comparator, in FNN stroke of the working output is connected to EB and input of the threshold voltage, a second input connected to the output FNN comparator FNN second output connected to the base member multipoint input.

The design of the neural-like base element is new and constitutes a separate invention.

In one embodiment of the invention, the neural-like basic element described is grouped with many of the same basic elements made on a flat substrate, forming a modular block of basic elements having two one-sided or two-sided contact pads with inputs and outputs parallel to each other.

The third technical task is to create a way of organizing the work of an adaptive control device. Not only the claimed device, but also a number of others, including computer control systems, can work on the principles described below, and the claimed method itself constitutes an independent invention.

To solve the problem in a known way of organizing the operation of a control device, which includes receiving electrical signals from sensors, processing them with a neural network, including establishing time-varying electrical connections between many artificial neurons, each of which is able to receive and summarize electrical signals from nearby neurons or from external sensors, generate a new signal and transmit it to other neurons or the actuator of the control device using so the following procedure. A part of the mentioned connections is established before receiving signals from the sensors and they are used to train the neural network, a part of the electrical connections are set automatically during the operation of the control device, for which the total signal received at the input of this neuron depends on the totality of all the neural connections currently established network, compared with the threshold excitation potential, the value of which is made dependent on the frequency of previous excitations of the neuron, and if it is exceeded give a given neuron, namely, they generate a normalized negative impulse at the input of a neuron, at the same time or with a short delay relative to the duration of the mentioned impulse, form an identical positive impulse at the output of a neuron with the possibility of establishing between neurons whose excited state at least partially coincides in time of the said electrical connection, moreover, by means of the aforementioned negative impulse, incoming electrical signals are partially neutralized, ensuring competition of neurons for n signal irradiation and preventing a situation where the signal multiplication factor exceeds unity, due to which the conductive structure of the neural network is continuously modified, including the formation of many feedback loops between neurons, while the output control signals of the control device or some of them are used to actively influence the parameters environment and / or in order to correct the conditions for receiving electrical signals from sensors.

The described method may have a number of specific embodiments.

Communications between neurons can be of a chemical, capacitive, resistive nature. In the particular case of the method, it is preferable to carry out the aforementioned resistive connections, which can be realized, for example, by the breakdown of a dielectric separating neurons, or even based on conventional relay elements, if the dimensions of the device are allowed. Moreover, by analogy with the synapses of biological neurons, aging of resistive bonds can be realized, for example, by specifying the special properties of a dielectric film with a semiconductor layer or due to diffusion processes in a dielectric.

The delay of the front of the positive pulse at the output of the BE relative to the front of the negative pulse at the input of the BE is designed to increase the likelihood of the formation of links between events that reflect the situation "event A follows event B", which is more effective than the formation of the connection "before event B there was event A".

Learning a neural network can be the development of a given reaction to the test configuration of external signals, for which a neural network is pre-established between neurons from the sensor to the output of the control signal.

To develop the required response of the neural network to external training signals, some kind of behavior criterion is required (minimization of consumed energy, constancy of dissipated thermal energy, etc.), the deviation from which is corrected by training the neural network. In the particular case of the method, this essential feature is to minimize the number of received external signals. Note that for an adaptive control device and the method of organizing its work according to the invention, any external signals are educational, since the learning process continues without a teacher in the process of the device.

The process of training the neural network of the control device can be significantly accelerated if the method includes processes that simulate fatigue, restore working capacity and self-excitation of a biological neuron. Fatigue is provided by a stepwise increase in the threshold excitation potential in the case of frequent excitation of a neuron, when at each subsequent excitation the potential receives a pulsed addition along a circuit with low resistance. This causes the excitation of neighboring BEs and the passage of the signal through other electrical circuits in the absence of the correct reaction after repeated passage of the signal through this BE. The restoration of the base element is provided by a gradual decrease in the threshold potential, up to the reference potential, for which a high-resistance leakage circuit is provided.

Self-excitation is realized due to random potential fluctuations at the BE input. Such fluctuations, for example induced EMF, are rare and small in magnitude, therefore, the calculated reference voltage should be quite low.

The proposed method can be fully implemented by computer software. A control device with a relatively small number of BEs can be modeled on a conventional personal computer.

Thus, the adaptive control device contains the required number of sensors, a neural network of neural-like basic elements and the required number of actuators. Technologically, it is made of many of the same type of modular blocks, on each of which with an area of 3-4 cm ² and a thickness of 0.1 mm, about 10 ⁴ neural-like basic elements can be grouped, the design and operation of which are described in detail below.

The technical result is to simplify the design of an adaptive control device, expand its functionality and complicate the class of tasks.

The invention is illustrated by diagrams.

Figure 1 presents a schematic diagram of a control device.

Figure 2 schematically shows the design of a neural-like base element.

Figure 3 shows a diagram of a modular block of basic elements.

Figure 4 shows the installation of the control device from modular units.

Figure 5 shows the encoding process of the total input signal.

Figure 6 shows the correction of the stimulus-response chains when training a neural network.

Figure 7 shows a diagram of the simplest physical model equipped with a control device.

The terminology used in the description is often borrowed from neuroanatomy and neurophysiology. In this regard, all the analogies of the described neural-like elements and neural networks with their biological prototypes should be considered with a great deal of conventionality.

Description of the control device.

The control device (Fig. 1) contains neural-like basic elements 3 with inputs 1 and outputs 2. The inputs of some arbitrary BEs are connected to sensors (not shown) and receive signals from external sources, the outputs of some arbitrary BEs are connected to actuators (not shown) and give control signals. The basic elements form a multilayer structure and are located so that between the inputs and outputs of the adjacent BE separated by a dielectric film with a semiconductor layer, electrical connections a, b, c, d, etc. can be established. Communications conduct a signal in one direction, from the output of a given BE to the input of another BE close to it. The nature of such connections is shown on the remote element in figure 1. The bonds are formed on the basis of well-known technical solutions when a voltage pulse causes an electrical breakdown of the barrier layer of the film, which leads to the formation of conductive connecting channels between the discharge buses. In this case, a pn diode is turned on between the discharge buses (see, for example, Electronics: Science, Technology, Business 6/2001. M. Valentinova. Exotic memory; Pat. US 5835396, Three-dimensional read-only memory, 1998) .

According to the invention, each output of the BE can be connected with many inputs of other BEs, each input of the BE can be connected with many outputs of other BEs, the maximum number of connections depends on the particular topology of the multilayer neural network. In addition, numerous feedbacks can be formed between BEs, indicated by the number 4 in FIG. 1. Feedbacks can be short, as between elements k2 and j1, or long, as between elements k1 and i2. The diagram shows as an example only three groups of basic elements i, j, k with n elements in each and three feedbacks, but there can be many such groups and relationships according to the invention.

According to the invention, a part of the electrical connections is established during manufacture, for example by soldering, and is used to perform some basic reactions by the device when teaching a neural network with a teacher. The same connections are necessary for the device when learning without a teacher, i.e. during work.

For example, input i1 is connected to a photocell, and output k2 is connected to an electric motor moving this photocell. Pre-established communications form an electric circuit that turns on this motor when a photo signal is applied to the sensor.

The control device operates as follows.

Arbitrary inputs of the control device receive signals from sensors that convert dissimilar external stimulus signals (optical, acoustic, temperature, pressure, electromagnetic, radiation, etc.) into electrical signals of standard amplitude and duration. The design of the neural-like BE is such that the signal strength is automatically encoded by the number and frequency of pulses, as shown in FIG. At the moment of crossing the threshold voltage line with the signal line, the base element is excited and generates a normalized pulse at the output, and when the signal is large, a series of pulses, as shown in the lower line of the diagram in Fig. 5. Each such impulse can contribute to the total potential at the set of inputs of those basic elements with which the corresponding electrical connections are established. With each BE excitation, the threshold voltage increases stepwise. It can be seen that in the initial part of the diagram, the BE transmits even small signals, but at the end of the diagram, the BE does not pass signals of a much larger magnitude. This means that the signal can get to the inputs of several basic elements, excite them and lay new connections in the neural network. In addition, choosing a possible path through communications, bifurcating and decreasing in magnitude, the signal can return to the input of the previous base element, making a feasible contribution to the total potential of other signals arriving at this input.

The set of input signals from the sensors (stimulus) passes through the tree-like set of electrical circuits that has developed in the UE at a given point in time (see Fig. 6a), creating all new connections between the basic elements. The neural network not only processes the signals, the signals themselves continuously modify the conductive structure of the neural network. Close in location and simultaneously excited BEs leave a memory mark in the signal path - an electrical connection, so that the path of a similar signal configuration is “remembered”, and in the future, the neural network performs a predictive function. Ultimately, the signals are fed to the outputs of the control device, connected to the actuators, and become control signals.

The neural network topology described below provides that as connections grow in the control device, connections are formed between the outputs of the basic elements, logically located further from the input of the sensors, with the inputs of the elements logically located closer to the input. This implements local feedback, which is involved in the formation of the reaction. With large amounts of UE, local feedback circuits provide the possibility of long-term circulation of the internal configuration of signals along the structures of the neural network with sequential excitation of many neurons of the circuit, including feedback. So the number of sequentially excited BEs in the signal chain is obviously much larger than on the direct path from the sensor to the actuator. On the one hand, the presence of such internal connections allows a more adequate response of the device to the configuration of the input signals. On the other hand, the circulation of the signal realizes the forecast of the sequence of events, since a similar configuration of signals has already passed through this circuit and for new signals arriving at the input of the BE, there is already some addition to the total potential, we can say, “stand”, with which it is easier to exceed the threshold voltage.

Consider the structure and operation of a neural-like basic element.

A diagram of one of the basic elements, the combination of which ensures the operation of an adaptive control device, is shown in figure 2. The basic element (BE) 3 contains an input 1, an output 2, and a signal conversion unit, including a comparator 6, a threshold voltage driver 7 and two normalized voltage drivers, an FNN 8 and an FNN 9. The input of the threshold voltage driver 7 is connected to a reference voltage source. The first input of the comparator 6 is connected to the input 1, the second input of the comparator 6 is connected to the output of the threshold voltage driver 7, the output of the comparator 6 is connected to the input of the voltage isolation device 8 and to the input of the voltage isolation voltage 9. For the BE to work, an external electric power supply is required: an adjustable reference voltage source, a positive source potential, source of negative potential.

For the manufacture of a compact multilayer neural network, the basic element can be structurally grouped with many N of the same basic elements, the combination of which forms a modular block of basic elements made on a flat substrate, as shown in Fig.3. The inputs 1 and outputs 2 of the above-mentioned set of basic elements are located parallel to each other in the same plane, forming two contact pads. Both inputs and outputs can be laid on both sides of the substrate.

Neuro-base element (figure 2) works as follows. The input 1 from neighboring BEs or from external sensors can be fed with a variety of signals S _i , which are summed in accordance with the currently established multiple connections 5. The sum of the input signals is fed to the first input of the comparator 6, while the output of the threshold voltage generator 7 to the second input of the comparator 6 serves some threshold potential. The threshold potential is formed individually in each BE based on the reference potential applied to the input of the threshold voltage driver 7. At the moment when the sum of the signals at input 1 exceeds the threshold potential, the BE is excited. When the base element is excited, the comparator 6 instructs the generation of normalized voltages and the voltage collector 8 generates at the output 2 a positive potential normalized in magnitude and duration, which determines the duration of BE excitation and is an output signal of a neural-like basic element. At the same time, FNN 9 forms a negative potential at input 1, which is equal in magnitude and duration to the positive potential described above.

A similar process proceeds on the basic elements of the next section (see figure 1). At some point, two basic elements in neighboring blocks simultaneously find themselves in an excited state, for example, voltage + U is formed at output 2 of element i1, and voltage -U is formed at input 1 of element j2. Since the output of the excited element i1 and the input of the excited element j2 are separated by a dielectric with a calculated breakdown potential Ubreakdown <2U, a breakdown of the dielectric layer occurs and a resistor bond is formed between the base elements, indicated by a on the circuit. If only one BE is excited, communication is not possible.

The output 2 BE 3 can establish electrical connections with many neighboring BE, as shown in the diagram of figure 2 as a set of signals S _i .

The negative potential generated when the BE is excited through a low-resistance resistor (not shown in FIG. 2) is fed to input 1, sharply decreasing the total potential of the input signals S _i during the BE excitation period, thereby ensuring competition of the basic elements.

With frequent excitation of BEs, a positive potential is applied to the accumulator capacity of the input of the threshold voltage driver 7 (not shown) from the output of the low voltage filter 8 and the threshold voltage increases so much that the BE ceases to be excited, simulating the synapse fatigue of a biological neuron. A negative pulse ceases to be supplied to the input 1 of the BE and ceases to “sit down” the total potential of the signals S _i . This contributes to the excitation of other BE adjacent to the links with tired BE. With a long absence of excitations, the threshold potential flows through a high-resistance resistor (not shown) and slowly decreases, tending to the reference potential Uoporn. In this way, the recovery of BE performance is realized.

According to the invention, the basic element of the neural network, having a multi-point input, operates not with environmental signals, but with a superposition of signals received from different sources. Since the signals are transmitted from the sensors to the actuators only through excited basic elements, the signal received at the input can be transmitted in two cases: either it has a value greater than the threshold, or several signals have arrived at the input and their total value is higher than the threshold, i.e. . BE excitation condition is satisfied. When excited, a new signal is generated that can be transmitted over the neural network. The signal is transmitted either through a newly established connection, or through communications established earlier. If there are no connections, the excited BE does not transmit a signal further. Since the signal is transmitted from layer to layer of the neural network not in its original form, but is constantly generated, the signal does not fade.

The resistance of the newly formed bond R (t ₀ ) = Rmin increases with time R (t> t ₀ )> R (t ₀ ). In Fig. 1, the connections b and c at the input of element j1 have different conductivities, so that two signals of the same magnitude from the outputs of elements i1 and in make a different contribution to the total signal fed to the input of the comparator of element j1, which eliminates the synapses that calculate the weight of the signals in the prototype device. In the device according to the invention, competition of the basic elements for the signals arriving at their inputs contributes to the fact that for the current conductivity of the bonds, one of the ECs has an advantage and determines the path for the signal to follow as the most suitable for this configuration of input signals.

A multilayer neural network is mounted from identical modular blocks MBi by superimposing pads on each other, orthogonally, as shown in Fig. 4, or at some angle to each other. In the latter case, the architectural options of the neural network can be significantly expanded. The assembly can be made in the form of spirals, wells, hexagonal cells, their combinations, etc. The number of layers is practically unlimited and depends on the complexity of the control tasks that the adaptive control device will solve. The planes of all blocks remain parallel. Orthogonal to the planes of the blocks, power supply lines Uoporn, U +, U- are laid, as shown in Fig. 3, which are powered from sources common for assembly. Unlike traditional neural networks, in which layers of neurons follow each other, the invention provides for the contact of the outputs of subsequent modular units with the inputs of the previous ones, which ensures the formation of feedbacks. In the case of complicating the task, it becomes possible to increase the number of layers and the number of modular blocks in each layer during the operation of the system.

BE electronic elements operate at low currents. If necessary, BEs can always be equipped with means for converting them into signals of the required power.

The organization of the formation of potentials at the output and at the input of the BE described above allows us to solve two main problems:

a) ensure the establishment of an electrical connection between two simultaneously excited BEs by means of a breakdown trace of a dielectric having a specially selected specific electric strength and separating the output of one BE from the input of a neighboring BE. The modular condition for establishing a new relationship between BEi and BEj:

| Uout i-Uin j |> | Udown | with | U breakdown |> | Uout |, | Uin |

b) for any configuration of the input signals, ensure that only the BE is excited, which, due to the increased weight of the signals, has the highest value of the sum of the potentials at the input. This implements the competition of basic elements that are associated with this configuration of input signals: the excitation of the first BE prevents the excitation of other BE, the inputs of which receive signals that participated in the excitation of the first excited BE.

With such an organization of UE, many new relationships are continuously formed, but UU does not self-excite. The average number of BEs excited in the next cascade is no more than the number of exciting basic elements of the previous cascade, so self-excitation of the CI is impossible. In other words, in the control device, the multiplication factor of the signal K does not exceed unity. When m signals are fed to the input of the first stage, the output of the Nth stage will contain on average m ₁ = m · K ^N signals and for K≤1 always m ₁ ≤m.

Thus, when implementing the method according to the invention, signals from sensors converting dissimilar external stimulus signals (optical, acoustic, temperature, pressure, electromagnetic, radiation, etc.) into electrical signals of standard amplitude and duration, the total signal at the input of the BE, they are compared with the threshold excitation potential, and the strength of the stimulus signal is encoded by the number and frequency of normalized pulses.

The set of input signals from the sensors, passing through the branched set of electrical circuits of neural-like elements that has developed in the UE at this time, including short and long feedback circuits of the neural network, is processed. Ultimately, the signals are in the form of a multi-dimensional response, i.e. a modified set of normalized signals, get to the outputs of the neural network connected to the actuators of the control device. According to the method according to the invention, the control signals of the device or part of them are used to actively influence the parameters of the external environment or to correct the conditions for receiving signals from sensors. In particular, the control signals provide the actions of the actuators aimed at reducing the input signals, i.e. carry out global feedback, not necessarily electrical, between the input and output of the control device through the external environment.

The action of actuators to reduce input signals can be very diverse, from moving the control unit in space to eliminating signal sources, which leads to a change in the entire configuration of external signals entering the inputs of the control unit.

Simple (basic) reactions to elementary signals of sensors are determined by laying a certain number of connections between the basic elements in the manufacture of the device. Complex reactions form the primary learning of the neural network, after which the device adapts itself.

The configuration of the input signals in the initial state is a function of the environment, the number and location of the sensors. The passage of signals through a conductive medium of the control unit leads not only to a visible reaction of the control unit to input signals, which manifests itself in the action of actuators, but also to a continuous change in the internal conductive structure of the control unit under the influence of transmitted signals. This change occurs due to the formation of new, initially most significant electrical bonds between pairs of simultaneously excited neighboring BEs.

According to the invention, over time, the weights of the bonds formed between BEs decrease, which models forgetting, aging of the bonds of biological neurons. In the general case, aging of bonds in various places of a neural network can occur at different rates. The incoming signals are able to excite other BEs and pass through the UE through new electrical circuits. A change in the structure of connections in the UE is manifested in a change in the reaction of the UE to the same configuration of input signals, which indicates a retraining of the neural network.

Note that the UU does not generate signals on its own. It only transmits signals through its structures with a reproduction coefficient at each stage K≤1. This leads to the fact that, in the absence of external signals, the control unit does not give any commands to its actuators, as if freezes until some disturbing signals appear at its input.

This behavior of the device is determined by the basic relationships laid during its manufacture: the device responds to the sensor signal with a completely unambiguous command to the executive mechanism, which reduces the value of the input signal. It is clear that the functions of the actuator provided by the tasks of the control unit must correspond to the nature of the signal of the external environment. The training of the device is based on the same desire of the control device to reduce the value of the input signal. If, as a result of its previous reactions to the input signals, the UE falls into a position with the absence of input (exciting, annoying) signals, it ceases to act. Note that this state of the UE is stable, similar to the minimum potential energy of the ball on an uneven surface, in contrast to the state when receiving sensor signals that prompt the device to use actuators, and is consistent with the principle of least action for mechanical systems.

When striving to reduce to a minimum one signal, the device may encounter the growth of another signal. As a result, it proceeds to the process of minimizing the sum of two signals, in the general case, the sum of all input signals. In other words, the UE strives to achieve its local goal - to get away from irritant signals, regardless of the degree of adequacy of the UU reactions to external signals, starting from the basic reactions specified in the manufacture of the UU.

This bears some analogy with the unconditioned reflexes of biological organisms. All unconditioned reflexes of a biological organism realize its desire to escape from danger, most often by changing its location. So, a simple biological organism tries to reduce irritation to zero, a more complex organism can carry out other, more complex cause-and-effect reactions.

The use of the principle of minimizing input (annoying) sensor signals for practical purposes can be illustrated by the actions of the control device for an elevator in a multi-storey building. Calling an elevator by a passenger should take precedence over a command from a control device that sends the elevator at a given time to a given floor. If the order of calling the elevator to the floors obeys some regularity, for example, in the morning at t1 - floor n, at t2 - floor n + 2, in t3 - floor n-2, etc., then you can optimize energy consumption by training the neural network to minimize the time of empty (without passengers) elevator trips between floors, for this the time of empty trips should be assigned as input (annoying) signals.

Very rare cases of spontaneous excitation of BEs with the generation of a signal that is not initiated by input signals can only remove the UE from the state of rest (equilibrium) for a short time, therefore these spontaneously generated pulses cannot significantly affect the behavior of the UE. Although they can (and are) lead to a change in the structure of the UE by establishing new relationships between self-excited and other excited BEs, in particular, to induce the device to get out of the situation where there are no signals from the sensors and repeat the entire cycle of responses to the changed configuration of the input signals.

Consider the process of primary learning adaptive control device (see figure 1).

Upon completion of the assembly of the control unit, the basic connections are established between some BEs, so that chains are formed from sensors to actuators, the inclusion of which leads to such a change in the configuration of the input signals that the input signals are reduced (leaving, moving from exciting signals without changing their source; “shielding "From exposure, etc.). The more such pre-established relationships are laid in the UE, the greater the number of basic reactions is provided in its design. After completing the factory settings, the control unit is ready for operation in the simplest situations. Pre-installation of basic relationships at the stage of manufacturing a CI determines the response of the CU to “dangers” and bears some analogy with the formation of unconditioned reflexes in a biological organism.

The factory setting of basic connections is carried out in hardware as follows:

a) a signal is applied to the sensor. At the same time, it is set to which BE (or modular unit) the signal is supplied and at the output of which BE the signal appears. It is determined which block the signal of the first stage comes from and to which blocks of the second stage the signal from the block of the first level is suitable. Between BE can be a connection. We do not lay it by hand, but create the conditions for its formation. These conditions are created as follows:

b) slowly lower the adjustable reference voltage at the second-level block so that self-excitation occurs in any BE from this second-level block and, as a result, the first-level BE forms a connection with the BE from the second-level block during its self-excitation. Note that prior to the formation of this basic connection, all BEs from the block of the 2nd level had no connections and could not be excited through any of their inputs;

c) the fact of the self-excitation of some BE of the 2nd level can be established by increasing the current in the power supply circuits of unit 2. The fact of establishing a connection between the excited BE of the 1st level and one of the self-excited BE of the 2nd level can be established, for example, by radio emission from the contact area at the time of breakdown. It is advisable to connect the ends of the input-output conductors of the entire block and draw a conclusion for monitoring, and after establishing the basic connections, remove the installation conductors;

d) return the reference voltage on the block of the 2nd level to normal and slowly lower the reference voltage for the block of the 3rd level, achieving self-excitation of some BEs in its composition, and, continuing to give signals to the input of the BE of the block of the 1st level, carry out the procedure, similar to b) and c), to establish a basic connection that conducts a signal between the output of the excited BE of the 2nd level and the input of the self-excited BE of the 3rd level;

e) repeat a similar procedure for laying the basic connections from the sensor to establish communication with the drive of the desired actuator, choosing the desired signal path by selecting a block at each stage of laying the next link of the installed circuit. So, upon completion of the communication installation and return of the reference voltages for all units to normal, when the signal is input to the sensor input, we obtain the desired “manually” set reaction at the output of the control unit in the form of an actuator action that reduces this input signal;

f) the procedure for establishing basic relationships from the initiating sensor to the reaction that reduces its signals is repeated the required number of times until the necessary set of initial reactions is created in the UE, providing the possibility of further learning and functioning of the UE.

Pre-installation of basic connections is necessary, but not sufficient for the operation of a multifunctional UU.

Adaptable devices should learn from their own experience, and at the initial stage - through primary training, receiving input signals from the instructor. As soon as the device creates a model of its reactions, it will be able to recognize analogies based on past experience to predict future events and propose solutions to new problems for it. There is a need to create ways of high-speed learning control device.

After completion of the formation in the UE of an acceptable number of basic reactions, primary training is carried out, expanding the set of adequate (desired) reactions to the configuration of the input signals of the sensors.

Primary training is based on the artificial suppression and natural forgetting of electrical connections that give incorrect control signals, and the organization of new connections between neural-like basic elements, leading to increasingly adequate responses of the control device to incoming sensory signals. The neural network is trained by replacing links that give the wrong reaction, with new connections that give more and more correct responses. Thereby, behavior patterns are set in the control device for generating the correct reactions, after which the device adapts to the changing external signals independently.

Neural network training is carried out by applying to the device corrective signals presented by the teacher (see Fig. 5 and 6). The reaction of the UE to external signals consists in turning on the actuators that received control signals after the sensor signals passed through pre-established and existing new neural network circuits, from the sensor signal to the reaction, as shown in Fig. 6a. Here, BR is the basic reaction (an analogue of the conditioned reflex), HP is the wrong reaction, and PR is the correct reaction. If, as a result of the UR reaction, the signal-sample decreases, this is the correct reaction (PR). Otherwise, we observe an incorrect reaction (HP). If the signal strength of the sensor increases, the response to a series of pulses arriving at the output is also enhanced, both due to the more frequent passage of the signal along the same circuit, and due to the propagation of the signal along parallel circuits, which is embedded in the design of the neural-like element.

Primary education is conducted with the teacher according to the following scheme.

A sample signal is given. If the response of the UE to the presented signal-sample is correct, then the memory of the UU remains the newly executed chain of links, which will be reproduced upon subsequent presentations of this signal. If the reaction of UU to the sample is incorrect (HP), then the teacher (instructor) includes a correction signal — a strong stimulus signal (bottom diagram in FIG. 6b), causing a large number of BR pulses to pass through the chains of basic reactions (thick lines in FIG. 6a) ) As a result of the gradual growth (branching) of bonds in the control unit, regions with electrical circuits appear along which both the signals of the incorrect response of the control unit to the sample and the reaction signals of the control unit to the correction signal pass. This allows the supply of a correction signal to cause BE competition and their fatigue and block the signal from passing through the incorrect reaction circuit.

As a result of fatigue and blocking of the incorrect reaction circuit, a new chain of reaction of the control unit to the sample signal is formed. If this new reaction chain also does not suit the teacher, he continues to provide a strong corrective signal to the input of the UE. When the UE response to the sample signal becomes correct (PR), i.e. the reaction of UE leads to the shutdown or decrease of the sample signal, the teacher turns off the correction signal. The training ends. With the subsequent presentation of the same sample signal, the control unit will act according to the latter option, since the communications that implemented it have the greatest weight.

With prolonged failure of the supply of the correction signal, the BR circuits grow, exciting a large number of nearby basic elements. Finally, they reach the region of passage of the HP chains to the sample signal (the upper line of the signal chain in Fig. 6a) and bring the result - they block the chains of the wrong reaction.

This is how simple reactions are taught using a strong correction signal. In this case, the efficiency, effectiveness of the correction signal is quite simple, however, the configuration of the sample signals must be selected so that the correct reaction is fast enough.

Learning a complex reaction is carried out stepwise, in several stages. In several stages, on the basis of a strong correction signal, the UE response to a much smaller correction signal is formed. The multi-stage training consists in the inclusion of increasingly stronger corrective signals in the absence of a correct UE response to the presentation of a weak corrective signal. To do this, before applying a strong correction signal, a weaker correction signal is turned on to create an electric circuit for predicting the possibility of applying a strong correction signal if the reaction does not change.

The operation of the adaptive control device in its simplest execution is illustrated by an example (Fig. 7).

The cart is equipped with a control unit (UU), a power source, an audible siren. There is a block of movement of the database, under the action of which the trolley can move left and right between two walls with a distance of 2 meters between them. There is a strong light source L from above, representing any additional external signal. A strong light source can be turned on by the experimenter and turned off automatically by a sound signal. With the help of a siren, the UU can give such a sound signal. The cart movement block (DB) models external influences, in the absence of which the cart is at rest (the neural network falls asleep). At random times, the database includes the movement of the cart at a random distance in the range of 0.4 ÷ 0.6 m in a random direction (left or right).

A block of BS sensors with sensors is fixed on the trolley:

sensor C1 contact with the left wall;

sensor C2 contact with the right wall;

sensor C3 dangerous proximity with the left wall at dist. 0.2 m (LL lamp);

dangerous proximity sensor C4 with the right wall at dist. 0.2 m (LR lamp);

high light sensor C5 from above (lamp L).

When receiving signals from sensors, the control unit can send three signals to the cart database: an order to move 0.3 m to the right, an order to move 0.3 m to the left, an order to give an audio signal. The signals from the control device have a higher priority than random movement signals from the database motion block.

The control unit is mounted, for example, from 10 BEs with 20 inputs in each so as to physically provide the possibility of the formation of connections between any BEs. Before testing, we carry out factory settings in the UU according to the algorithm described above, namely, we lay the chains of basic reactions.

The signal of the sensor C1 includes the BE of the contact with the left wall, which includes the BE "enable movement to the right", which gives the signal to the DB the order "to the right." This chain of connections ensures the functioning of the basic reaction “when turning on the left wall, turn on the movement to the right”.

The signal of the sensor C2 includes the BE of the contact with the right wall, which includes the BE "enable movement to the left", which sends a signal to the database (the order is "left"). This chain of connections ensures the functioning of the basic reaction that was laid in advance “when turning on the right wall, turn on the movement to the left”.

To input signals from sensors to the control unit:

connect the C3 sensor with the corresponding BE;

connect the C4 sensor with the corresponding BE;

connect the C5 sensor to the BE of the strong light sensor.

Initially, the basic elements are not involved in the reaction chains of the UE to signals.

a) Adaptation (self-training) of the neural network.

Turn off the output of signals for all sensors, except for contact sensors with walls. The behavior of the SU will be very simple: during random walks from random signals from the database, upon contact with one of the walls, the trolley will move away from the wall.

Now we turn on the signals for the sensors of dangerous proximity with the wall at 0.2 m. In the case of the first operation of the proximity sensor with the right wall, but without contact with it, nothing will happen in the control unit. But if the proximity sensor works with the right wall and the wall is reached in the same maneuver to the right, then in addition to the “left” order established during the assembly of the order, two BEs will be simultaneously activated and the electrical connection will be established in the control unit “the right side follows contact with the right wall”. So the next time when the proximity sensor with the right wall is triggered through the above-mentioned connection from the BE output of the dangerous proximity to the contact input of the contact with the right wall, an electrical signal will arrive and the command for the database will appear “left” at its output, since simultaneously by the excitation BE of the approach, the BE of contact with the wall will also be excited. So, subsequently, “dangerous” contact with the right wall will never take place - the training of UU has occurred. In the same way, communication will automatically be established when interacting with the left wall and an acquired reaction to convergence will arise.

b) Retraining of the neural network.

If now the conditions for the interaction of the control unit with the external environment are substantially changed - to swap proximity sensors to the walls, then when approaching the right wall, when the former proximity sensor to the left wall is triggered, according to the existing connection “contact is coming from the left wall to the left” the car will turn on the movement to the right and will collide with the right wall. But upon reaching the right wall, the basic “unconditional connection” set during manufacture will work and the cart will receive an order “to the left”. In the same episode, there will be simultaneous excitation of the BE, which previously received a signal of approaching the left wall and the BE of contact with the right wall. A new connection is formed between them, stronger than the time that the connection between the BE of the approach to the left wall and the contact BE of the contact with the left wall has become obsolete. Therefore, with subsequent excitations of the former BE closer to the left wall, it will more quickly excite the BE of contact with the right wall and turn on the “left” signal-order. An excited BE of the contact with the right wall will immediately “feed” the voltage at the output of the BE of the former approach to the left wall, so that the BE of the contact with the left wall will no longer be excited and will not issue an order “to the right”. In the same way, the retraining of the SU for the BE of the former approach to the right wall will occur.

c) Search for a solution in an unknown situation.

We turn on the bright light, as a result of which the BE of the bright light is excited. At this moment, some BE motion can work, for example, proximity BE. The cart, under the influence of the erroneous connection between the bright light BE and the proximity BE, will begin to move, because it has no other reactions yet. As mentioned above, a stable state of the system is the absence of signals. In fact, the control device goes through possible reaction options, trying to turn off the upper light, i.e. reduce the number of signals. For some time - to no avail, the light will not turn off. The trolley reaches the wall (contact) and rolls in the opposite direction. And it will ride like that for a while, but the BE of contacts with the wall will begin to tire, their response thresholds will increase, so that it will be possible to self-excite the basic element for switching on the siren. The siren will turn on, a strong light will automatically turn off, which is inherent in the conditions of the task. So, as a result of a random search for UU, a way out of an almost deadlock situation will be found. The established relationship between the BE of the bright light and the BE of the siren is maintained (within the aging of communication). So in the next episode, a siren will sound to irritate with a bright light and the lamp L will be immediately turned off.

The described example illustrates the essential features that ensure the fatigue of the base element, self-excitation of the base element and the aging of its bonds. Indeed, each BE has a current threshold voltage for the comparator to operate, based on a given reference voltage. Frequent activation of the BE leads to an increase in the internal threshold voltage and to a loss of the ability to excite (the case with multiple left-right movements), a long stay of the BE in an unexcited state leads to its self-excitation as a result of lowering the threshold voltage below a certain limit (the case with the siren on), and a decrease in the conductivity of old connections changes the signal path (the siren signal is ahead of the motion signal). The technical parameters of the base element are determined by calculation. In the same way, it is not difficult to determine the duration of signals generated by BEs, the duration of motion pulses, and the reaction duration of the outputs of various sensors, which are calculated based on the physical conditions of the problem.

Information sources

1. Artificial Neural Networks: Concepts and Theory, IEEE Computer Society Press, 1992; J.A. Anderson, An introduction to neural networks, Chapter 2, MIT, 1999.

2. Patents RU 2128857, 1999; RU 2295769, 2007.

3. Patents RU 66831, 2007; RU 7208, 2008; RU 75247, 2008; RU 77483, 2008.

4. Pat. US 5347613, 1994; US 6643627, 2003; US 7426501, 2008.

5. Pat. US 5220641, 1993 (prototype).

6.http: //habrahabr.ru/blogs/hardware/14025/.

7. M. Valentinova. Electronics: Science, Technology, Business, 6/2001; Pat. US 5835396, 1998.

Claims (10)

1. An adaptive control device containing many neural-like basic elements (BEs), each of which has an input and an output, is able to receive and summarize electrical signals from sensors or from other BEs and enter an excited state with the generation of an electrical signal that can be transmitted to many other BEs or to the actuator of the device, grouped in M> 1 flat modular blocks with N> 1 BEs in each so that the outputs of the BEs of one block are located orthogonally or at some angle to the inputs of the BEs block, forming at least one grid with N ² nodes, characterized in that the outputs of the BE of one block are superimposed on the inputs of the BE of another block through a dielectric film with a semiconductor layer, each of these grid nodes when creating a potential difference sufficient for breakdown dielectric, has the ability to form a resistive connection between the output of one BE and the input of another similar BE, the conductivity of which decreases with time, the inputs of some arbitrary BEs are connected to sensors and are assigned to receive signals from external sources, the outputs of some arbitrary BEs are connected to the actuators of the control device and assigned to issue control signals, some of the resistive connections being laid beforehand in the manufacture of the control device and, when the signal is supplied, forms at least one electrical circuit from the input connected to the sensor to the output connected to the actuator. 2. A neural-like basic element of an adaptive control device, comprising a multi-point input for receiving and summing electric signals from similar basic elements (BE) or from external sources, a working output and a signal processing unit capable of generating electrical signals upon excitation of a BE, characterized in that that the signal processing unit comprises a threshold voltage driver connected to a source of adjustable reference voltage, a comparator and two normalized voltage drivers (LVF), the first of which generates positive rectangular pulses, the second one generates negative rectangular pulses of the same magnitude and duration, the comparator is connected to the multipoint input of the BE and with the output of the threshold voltage shaper, the input of the first TNF is connected to the output of the comparator, the output of the first TNF is connected with the operational output of the BE and with the input of the threshold voltage driver, the input of the second voltage converter is connected to the output of the comparator, the output of the second voltage converter is connected to the multipoint input of the base element. 3. The neural-like basic element according to claim 2, characterized in that it is grouped with many of the same basic elements made on a flat substrate, forming a modular block of basic elements having two one-sided or two-sided contact pads with inputs and outputs parallel to each other. 4. A method of organizing the operation of an adaptive control device, including receiving electrical signals from sensors, processing them with a neural network, including establishing time-varying electrical connections between a number of artificial neurons, each of which is able to receive and summarize electrical signals from nearby neurons or from external sensors, generate a new signal and transmit it to other neurons or the actuator of the control device, characterized in that part of ides are installed before signals from sensors are used and they are used to train the neural network, some of the electrical connections are established automatically during the operation of the control device, for which the total signal received at the input of this neuron is compared with a threshold excitation potential, the value of which is dependent on the frequency of previous excitations neuron, and if it is exceeded, the given neuron is excited, namely, they form a normalized negative impulse at the input of the neuron, simultaneously or with little with respect to the duration of the mentioned pulse, an identical positive pulse is formed at the output of the neuron with a delay, with the possibility of establishing between the neurons, the excited state of which at least partially coincides in time, of the aforementioned electrical connection, and by means of the said negative pulse, the electrical signals arriving at the input of the neuron are partially neutralized, ensuring neurons compete for receiving a signal and obstructing a situation in which the signal multiplication factor exceeds unity tsu, due to which they continuously modify the conductive structure of the neural network, including the formation of many feedback loops between neurons, while the output control signals of the control device or some of them are used to actively influence the parameters of the external environment and / or to correct the conditions for receiving electrical signals from sensors. 5. The method according to claim 4, characterized in that said electrical connections establish a predominantly resistive type. 6. The method according to claim 5, characterized in that they provide a decrease in time of the electrical conductivity of the installed resistive connections. 7. The method according to claim 4, characterized in that the aforementioned primary training of the neural network forms a set of required output control signals for a given configuration of sensor signals. 8. The method according to claim 4, characterized in that the said primary training of the neural network is taught the control device to minimize the number of input signals. 9. The method according to claim 4, characterized in that said threshold excitation potential is set equal to a certain reference voltage, provides an increase in the threshold potential with an increase in the number of excitations per unit time and its gradual decrease up to the reference voltage with a decrease in the number of excitations per unit time or at their absence. 10. The method according to claim 4, characterized in that they provide self-excitation of the base element at a low level of the threshold potential due to random fluctuations of the potential at its input with the possibility of the formation of new connections between the basic elements in the absence of environmental signals.

Images