Continuously Evolving Programs in Genetic Programming Using Gradient Descent

This paper describes an approach to the use of gradient descent search in genetic programming for continuously evolving genetic programs for object classification problems. An inclusion factor is introduced to each node in a genetic program and gradient descent search is applied to the inclusion factors. Three new on-zero operators and two new continuous genetic operators are developed for evolution. This approach is examined and compared with a basic GP approach on three object classification problems of varying difficulty. The results suggest that the new approach can evolve genetic programs continuously. The new method which uses the standard genetic operators and gradient descent applied to the inclusion factors substantially outperforms the basic GP approach which uses the standard genetic operators but does not use the gradient descent and inclusion factors. However, the new method with the continuous operators and the gradient descent on inclusion factors decreases the performance on all the problems.