Средства демонстрации эволюционных методов настройки весовых коэффициентов искусственных нейронных сетей

Abstract

The problems of developing neurocontrollers for controlling dynamic objects are described, including the complexity of generating training data sets. It is indicated that one of the known methods of training an artificial neural network to control an autonomous driving agent is the neuroevolutionary approach, which involves the use of a genetic algorithm to adjust the synaptic weighting coefficients of the artificial neural network. The idea of using a tool for demonstrating the evolutionary approach to setting the weighting coefficients of an artificial neural network for practical training of students in the basics of the neuroevolutionary approach is proposed. The general structure of the genetic algorithm as a multicriteria optimization method is described. Basic information about the structure, principle of operation and methods of training multilayer artificial neural networks with the direct direction of information propagation is provided. Software has been developed to demonstrate the neuroevolutionary approach using the example of the evolution of an artificial neural network of a given structure designed to control a simplified computer model of an autonomous vehicle. The problems of empirical selection of the selection operator and the crossover operator when using evolutionary methods for training artificial neural networks are described. The known negative consequences of rearranging the values of the weighting coefficients of interneuron connections in the interval between two successive layers of artificial neural networks when implementing the crossover operator for artificial neural networks are indicated. The situation of stagnation of the evolutionary process, characteristic of a genetic algorithm, when the function being optimized reaches a local extremum in the process of searching for an optimal solution to the problem is described. A method for resolving the problem of stagnation when using an evolutionary approach for training an artificial neural network is described. A comparison was made of the effectiveness of using proportional selection and tournament selection during the trial operation of the developed software to demonstrate the neuroevolutionary approach using the example of the evolution of an artificial neural network of a given structure designed to control a simplified computer model of an autonomous vehicle. Options for using the developed software when teaching students the basics of artificial intelligence technologies and evolutionary methods of multi-criteria optimization are proposed.