Evolving Neural Network Architecture and Weights Using An Evolutionary Algorithm
Tania Binos · 2003
Artificial neural networks have been applied to a variety of classification and learning tasks. The success of error correction training algorithms such as backpropagation has meant that supervised learning, where the correct outcome is known, has been the most used. In addition, although the structure of an artificial neural network is a significant contributing factor to its performance, the structure is generally heuristically chosen. The use of evolutionary algorithms as search techniques has allowed different properties of artificial neural networks to be evolved. Several successful neural/evolutionary approaches such as the EPNet [35] algorithm are hybrid algorithms combining gradient descent methods such as backpropagation learning for connection weight training with evolutionary structure search. Generally the fitness measure used in these approaches is a mean or sum squared network output error over input training patterns. This type of fitness measure is applicable to supervised learning, but not necessarily to reinforcement learning tasks where the correct output is not generally known. Evolutionary search algorithms allow greater flexibility in the choice of a suitable fitness measure. As a result, neural/evolutionary algorithms that are not reliant on gradient descent techniques and not reliant on the standard fitness measure may have wider applicability outside the supervised learning domain. The goal