ADAPTATION OF NEURAL NETWORKS USING GENETIC ALGORITHMS

This paper proposes an approach to the problem of adaptation of neural networks (NN) to arbitrary tasks. It is well known that the functional properties of a NN depend on its construction: on topological structure, learning and activation methods, and signal output. A definition language is developed for describing Various constructions of NNs in the shape of strings. This paper uses a model of a neuron which has a receptive field and adaptable learning, activation and signaling, while the NN model consists of interconnected layers allowing feedforward, feedback and lateral connections with a single input and ouput layer. Adaptation of NNs is done with a genetic algorithm (GA) using crossover, mutation, and proportional selection operators on a population of strings that represent NNs. These strings (and their respective NNs) are evolved until they fmd solutions to given tasks which are defined as objective functions. The paper proposes a solution to >>deception<<, an important problem concerning GA's convergence: a strict hierarchy in the description of NNs based on ordered expression which decreases the probability of dual representations. This approach can develop autodidactive NNs.