V.M. Kureichik, B.K. Lebedev, E.V. Nuzhnov^{1} Neural networks instruction by means of genetic search methods
In order to the neural network (NN) receive the possibility to solve given task (if input signal is X, then output signal will be Y, as it's nesessary) they have to adaptate the network parametres. The adaptation produces on the base of educational examples sample, which consists of some pairs (, ).
From the mathematical point of view the NN instruction presents a non-linear optimizatio n task with many parametres. We may choose two large groups of algorithms: gradient and stochastic algorithms. Gradient algorithms of networks instruction are based on the partial derivative calculation of error function by all network parametres. The stochastic algorithms search error function minimum randomly. The most effective algorithm from the first group is the algorithm of reverse error performing. Among algorithms of the second group the most well known are algorithms which simulate the native process flowing. They are: parametric and alternative adaptation algorithms, genetic adaptation methods etc. Genetic evolution processes are developed on the base of analogy with brain models. It realize some brain exceptions which appeared as the biological revolution result.
In genetic approach the adaptation process for NN is considered as NN operation efficiency maximization or as error function D minimization. The evolution adaptation process structure is shown in fig.1, where IS – instruction sample, NN – neural network, GA – genetic algorithm, CA - control action, producing by GA, which creates the NN.
The GA development includes 3 base stages: principles of chromosome coding/decoding, main genetic operators, main structure and genetic search process. For fixed NN structure each chromosome may be presented as a vector H=(W,B), which stores values of semantic weights (W) and removals (B). The modified GA structure is described below.
1. The initial population formation.
2. Chromosomes evaluation in population and fitness function formation.
3. Chromosome pairs choice from the population.
4. Crossover operator application with probability P_{c}. If all pairs are analised, then go to 5, else go to 6.
5. Mutation operator application for each new chromosome with probability Р_{m}.
7. Segregation operator application for each new chromosome with probability P_{s}.
8. The end of algorithm.
O. Pyatkovskiy, D. Rubtsov, S. Butakov The Building of Information System with the Usage of "If - Then" Rules and Neural Networks
This work is devoted to an intellectual system for financial analysis in enterprise management. An advanced model of information technology of execution of financial analysis procedures is offered. The model allows to: 1)consider the forecast of financial indexes in the analysis of their current values; 2)perform additional learning of the system in its life cycle. The system of financial analysis has been made up with the usage of: 1)an expert system based on "IF-THEN" rules; 2)a neural network simulator, designed for formation of artificial neural networks. In the work the features of a neural network simulator and results of experimental usage of the system are indicated.
In this work we offer the model of the analytical system for elimination of some drawbacks in computer systems for financial analysis. Information processing at element of information system can be presented as the following transformation: , where – estimation of financial state of enterprise at a given moment , - information about financial operation of enterprise, - the forecast of for the moment . The use of the offered information processing method in the evaluation of financial indexes is intuitively clear - a manager will not give good evaluation, knowing in advance, that the situation will get worse after some time.
For the formation of the financial analysis system on the basis of the considered approach it is offered to use the following methods: "IF–THEN" rules and artificial neural networks. The expert systems based on "IF–THEN" rules are used to generate the evaluation of financial situation at upper level of hierarchy. The neural network simulator is used to form self-learning functional elements at all levels of hierarchy.
To evaluate the effectiveness of the usage of a neural network simulator for analysis of financial indexes the experiment on evaluating of a complex indicator of liquidity was conducted.
D.V. Andreev THE NEURAL PROCESSOR ON RESISTIVE RELATORS
The typical peculiarity of modern computing and cybernetic engineering is the high-performance processing of the continuous (analogue) information in real-time mode. The indicated feature to the full have the logical processors built in element base of relators [1] (analogue neural elements) and intended for the solution of standard problems of rank processing of analog signals (address or rank identification, selection, sorting and etc.). In the book [1] are reviewed potential and current relators, in which one switched values are accordingly voltage and current.
In a fig. the scheme neural processor, built in element base resistive relators [2] is shown, in which one switched value is the resistance. Each of these relators contains a differential comparator of voltage C, resistors R, closing S and opening analog switches, and at give on a control input of keys of logical unit the key S is selfcontained, the key is disconnected, and at give of logical zero point we have a return picture. In the scheme (fig.) relators are clustered in a matrix from lines and columns so, that k-th line contain relators. The resistance of i-th switching channel in offered neural processor is determined by expression
,
where ; - single function, - analog signals of input tuple . Everyone of a component of this tuple is characterized by the address i (sequence number in a tuple) and rank
With the registration (1) we shall receive:
.
Thus, resistance of i-th switching channel in the neural processor (fig) is proportionally to rank of a signal in a tuple .
If in the offered scheme follow-up to enter of n resistors with identical resistance R and of n operational amplifiers and to envelop everyone i-th operational amplifier by a negative feedback so, that its transmission factor has accepted a kind , the voltage on its output will be determined by expression , where E there is a given quantum of voltage on not inverting inputs of amplifiers.
In summary we shall mark, that on the basis of reviewed in the report neural processor the submachines of computing and cybernetic engineering oriented on address-rank processing of analog signals in real-time mode can be built. The advantage of offered neural processor is the regularity of its structure ensuring capability of its effective single-chip implementation on the basis of electron technologies of modern VLSI.
References
Volgin L.I. Synthesizing of devices for processing and conversion of the information in element base of relators.-Tallinn:Valgus,1989.
Andreev D.V., Sorokin A.V. To a problem on increase of response of devices of rank identification// Materials of international conference "Methods and means of conversion and processing of the analogue information".-Ulyanovsk:USTU,1999.-v.2-P.44.