Синтез ефективних стратегій управління технологічними комплексами харчових виробництв за допомогою нейромережевої ідентифікації параметрів нечітких когнітивних карт

Анотація

SYNTHESIS EFFECTIVE STRATEGIES OF TECHNOLOGICAL COMPLEXES THROUGH FOOD PRODUCTION NEURAL NETWORK PARAMETER IDENTIFICATION FUZZY COGNITIVE MAPS

N. Zaiets

The most effective tool for solving technological complex (TC) studies of the structure and obtain forecasts of its behavior under different control actions are fuzzy cognitive map (FCM). This neural information approach to optimization of cognitive maps allows the use of neural network learning algorithm that increases the efficiency of obtaining the appropriate models and creates preconditions for adaptation parameters in real time.

The sequence of tasks for the proposed approach can be defined as follows: identify basic objectives that characterize the studied process or system; identify the key elements of the process or system; detection of interference factors; construction of cognitive models; construction of possible scenarios of management; definition of criteria for evaluating scenarios; assessment scenarios and identify the best in terms of selected criteria.

Since most actual information on facility management obtained from expert estimates, it is largely subjective. And expert opinion on the same issue can significantly, sometimes fundamentally different. Therefore, the problem of optimal synthesis of expert opinions in order to build adequate FCM is relevant.

The aim - to create neural network unit of FCM synthesis support effective strategies for managing complex technology in order to optimize generalization of expert opinions when forming a matrix of values and concepts of its work in real time.

To achieve the goal formulated the following research objectives:

- Creating a structure for FCM management scenario TC;

- Development of methods of optimization of FCM TC using neural networks;

- Kohonen neural network synthesis support of FCM TC;

- Creation methods use Bayesian neural network to select coefficients FCM TC.

FCM task - to identify effective strategies and management scenarios given technological complex. This system is characterized by great uncertainty elements that belong to it (human, economic and other factors), and for the simulation of such a system can not get it accurate mathematical model. It is therefore advisable TC model represented as a generalized FCM. The appropriate structure fuzzy cognitive map was developed based on experimental studies and object-oriented analysis bakery.

The model of a TC is represented as a directed graph corresponding (fuzzy cognitive map), illustrating a plurality of links and nature of the interaction factors. Formation of values of weighting coefficients based on expert assessments must solve the problem of the impossibility of rapid survey of experts by changing the parameters of the TC. Therefore, experts (adopted three equal number of experts) assess not only the typical mode of operation of the complex process, but it is potentially possible value.

If the sequence proposed expert assessments are reduced to a single table with experimental data. In order to increase flexibility in evaluation each expert provides more than one possible meaning, and three most likely in its opinion mentioned weights. The data are grouped, and to determine a single value of weight coefficients based on expert opinion using neural network (NN).

The next step is forming adequate neural whose mission - calculation parameters FCM TC based on the information received on the subject. Spin software includes implementation of both modules provides adaptability condition: when the output of the system beyond the established effectiveness of possible conversion system.

To group peer reviews (clustering) and the definition of common values using self-organizing map. Kohonen model belongs to a class of vector encoding algorithms. It provides a topological display optimally placing a fixed number of input vectors in Higher dimension, providing thus compressing data .

After calculating using Kohonen self-organizing neural expert determined individual values of weighting coefficients FCM TC, with various combinations of input parameters, it is necessary to solve the problem of their choice in real time depending on the values of the sensor information and measuring complex. For this we use probabilistic neural network (PNN). This network does not require training in the sense required for the type of perceptron networks, radial basis function, etc., as all parameters PNN-network (number of elements and mentioned weights) are determined training data.

Using experimental data obtained and Kohonen neural network coefficients corresponding synthesized PNN-net. For each creates its own connection mizhkontseptualnoho probabilistic neural network, which provides optimal choice FCM coefficients in real time depending on the values of process parameters.

Using Kohonen neural network (optimization expert opinions) and Bayesian (selection coefficients) can synthesize efficient management strategies TC and improve functional efficiency FCM by providing opportunities to adapt in real time its parameters depending on the impact of various factors.