Artificial intellect in control systems of robots

Atanov S.K., the senior lecturer of chair «Computer systems» KazATU of S.Sejfullina

Technologies of an artificial intellect have always been closely connected with a robotics. Creation of robots - the cars, capable to operate as the person, is the general overall objective of these sciences. After the impressing successes reached in second half of the twentieth century at successful introduction of industrial robots in process of automated manufacture, now it is possible to speak about carrying over of the centre of scientific researches to area of creation of independent robots. Here it is necessary to mention space robots for studying of a surface of heavenly bodies of Solar system, robots for underwater researches. During struggle against terrorism there was a sharp necessity for the robots intended for mine clearing of suspicious subjects in places of a congestion of people. "Clever" robots which can extinguish fires without the aid of the operator are necessary, independently move on in advance unknown cross-country terrain, carry out rescue operations during acts of nature, technological failures, etc.

In modern understanding, the robot is the technical complex intended for performance of various movements and some intellectual functions of the person and possessing actuation mechanisms necessary for it, operating and information systems, and also means of the decision of vychislitelno-logic problems. Now it is possible to distinguish three classes of development of the automated systems:

· program - working under rigidly set program, a typical example of the MANAGEMENT information system technologically processes;

· adaptive - having possibility automatically to be reprogrammed to (adapt) depending on conditions, bases of the program of actions are initially set only;

· intellectual, here the task is entered in the general form, and the robot possesses possibility to make of the decision or to plan the actions in uncertain it uncertain or difficult conditions.

Theme of given article are features of decision-making in the second and the third cases. If is more exact, a choice of priorities at the analysis of signals of entrance signals. The increase in quantity of entrance signals from gauges leads to obvious complexities of training in case of the nejro-network approach. The same problems arise at the classical approach programming of microcontrollers owing to their low speed of processing of the information. At R independent entrance signals (gauges) dimension of the entrance alphabet of automatic model is defined as dim X =2 R. Parallelism introduction (as use of a neural network) does not rescue a situation - training time increases in a case after an exhibitor, also there are complexities of realisation of the given approach in microcontrollers in connection with their small resources of the RAM and ROM

The typical problem of classification of set of entrance signals or speaking automatically - recognition of situations, is resulted in drawing 1. Instead of an input-condition of transformation "input-exit" Y = R (X) presence of the additional device - the qualifier of conditions C is required. Qualifier C can be a various kind - from set produkt before realisation in the form of a neural network or the automatic approach. Its function consists in the analysis of an entrance vector and class definition which this vector concerns. The primary goal consists in creation of this qualifier as solving algorithm of adaptive behaviour of the robot.

Fig. 1. The converter "input-exit"

Let's consider mechanisms of reception of the adaptive qualifier on an example use of evolutionary modelling for reception of the qualifier in the form of the distinguishing automatic machine and application of the dynamic DSM-METHOD realising the qualifier in the form of set of rules.

Let's consider them on the classical test - movement on a black-and-white strip. The line is drawn on the field painted in chessboard order and is inverse. Strip gauges are formed by 4 steams the Ik-receiver/radiator as is shown in fig. 2.

Fig. 2 Arrangement of photogauges of definition of a strip

Classification of entrance signals can be carried out by means of DSM a method. The DSM-METHOD of automatic generation of hypotheses is the theory of the automated reasonings and way of representation of knowledge for the decision of problems of forecasting in the conditions of incompleteness of the information. Unlike classical DSM a method which works with the closed set of initial examples and their in advance certain properties, dynamic DSM the method allows to work in the open environment with quantity of examples unknown in advance.

The set of training examples is a set of pairs a kind

E = {ei} = {(X i, u i)},

Where Xi - a vector of signals of receptors,

ui - a management vector (a condition of executive mechanisms).

Elements of vectors of signals and management are represented by steams of binary values:

It is included = {01},

It is switched off = {10}.

Such representation is necessary for correct performance of operations of crossing and an investment over bit lines. On fig. 3 one of possible representations who has been used for movement training on a strip is presented.

Fig. 3 Structure of training examples and hypotheses

Hypotheses are represented in the form of set of pairs a kind:

G = {gi} = {{xi, yi}},

Where xi - a part of a vector of signals of receptors,

yi - a demanded vector of management (necessary action).

Dynamic ДСМ works in two modes:

- A training mode when there is a filling of base of the facts (set of training examples) and the hypotheses making the knowledge base are generated;

- Operating conditions when received before a hypothesis are used for development of signals of management.

In a mode of training for formation of training examples the external algorithm - so-called "teacher" is used. The given algorithm receives on an input the information from receptors and develops the operating signals necessary for adequate behaviour of the robot. Set of signals of receptors and the operating influences developed for them defines one training example. This example is checked on uniqueness and brought DSM by system in base of the facts. After entering of each new example in set of training examples search of hypotheses is made.

The received set of hypotheses {gi} will contain all possible crossings (the general parts) training examples. Further among them the minimum hypotheses, that is such which are put in the others are selected. Thereby the quantity of "useful" hypotheses is sharply reduced. The received minimum hypotheses are checked on uniqueness and brought in the knowledge base.

Training should be made until the knowledge base will not cease to replenish with new hypotheses. It is obvious, that in this case the training algorithm has touched all possible variants of entrance influences to which it is capable to react, and it is possible to consider, that the base of the facts {ei} is full enough.

In operating conditions DSM the system receives signals of receptors of which the test vector is formed on an input. Decision-making occurs by check of an investment of hypotheses in this vector. If in a test vector of signals of receptors the hypothesis the robot should operate according to it is put. If any hypothesis it is not found, this unknown condition for which it is necessary to generate a casual vector of management (or to do nothing, for example).

If the base of the facts is full, character of behaviour of the robot in operating conditions under control of DSM systems should differ nothing from management of "teacher".

On fig. 4 the fragment of the program of simple training algorithm for movement on a dark strip of range is presented. The given algorithm uses photogauges 2 and 3 for tracing of edge of a strip.

Fig. 4 Example of training algorithm for movement on a strip

Comparison of results of training

Both methods of inductive classification - on the basis of evolutionary modelling and a dynamic DSM-method - have appeared are applicable for the decision the robot of quite real problem - strip tracing. This problem demands for the realisation of small technical expenses, but is enough indicative.

Sufficiency of training examples. In the presence of representative sample of training examples both methods yield good results. However in the conditions of incompleteness of training set method ЭМ yields steadier results in comparison with DSM. It is connected first of all with character of management.

Consistency of training sample. DSM, unlike EM, it is not applicable in the conditions of contradictions in training examples. Such situation can arise, when the teacher is mistaken in an estimation of a condition of gauges. Errors of such type are necessary for eliminating at a stage of formation of training examples. In EM similar discrepancy not so is critical, since it leads at worst to uncertainty of behaviour.

Learning efficiency (speed). Training in EM - essentially long process. For steady training by a method of evolutionary modelling sometimes are required hundred thousand steps. In this respect the DSM-method possesses doubtless advantage - for training by means of DSM a method to receive some different training examples enough. In experiments to the robot was to pass one circle on real range enough that all necessary hypotheses were generated.

Dynamic training. Theoretically EM can work and in the open environment with quantity of examples unknown in advance, practically it is connected with the big computing expenses. Dynamic DSM the method allows to work effectively with in advance unknown quantity of examples at rather small computing expenses.

At realisation of practical algorithms there is a problem of limitation of computing resources of the independent robot. If modelling of evolution demands rather big time and capacitor expenses for work DSM of a method enough insignificant computing resources that allows to place the training and management program are direct on the robot.

The conclusion

If to accept as working definition, that as the intellectual robot is called the robot using for realisation of the algorithms of behaviour intellectual methods the created hardware-software device can be carried to a class of intellectual robots.

The literature

1. Tsetlin M. L. Researches under the theory of automatic machines and modelling of biological systems. TH.:Наука, 1969.

2. The grant on application of industrial robots. Under the editorship of Kazuhito Noda, transfer from the Japanese. TH.:Мир, 1975,-454 with.