Evolving Weights in Genetic Programs Using Gradient Descent


Authors: Will Smart, Mengjie Zhang
Source: GZipped PostScript (1068kb); Adobe PDF (305kb)

This paper describes an approach to the use of gradient descent search in tree based genetic programming for object recognition problems. To learn better partial programs, a weight parameter is introduced in each link between every two nodes in a program tree, so that a change of a weight corresponds to a change of the effect of the sub-program tree. Inside a particular generation, weight changes are learnt locally by gradient descent search, but the whole evolution process is still carried out across different generations globally by the genetic beam search. This approach is examined and compared with the basic genetic programming approach without gradient descent on three object classification problems of various difficulty. The results suggest that the new approach outperforms the basic approach on all problems.

[Up to Computer Science Technical Report Archive: Home Page]