Extending genetic programming to evolve perceptron-like learning programs

Abstract

We extend genetic programming (GP) with a local memory and vectorization to evolve simple, perceptron-like programs capable of learning by error correction. The local memory allows for a scalar value or vector to be stored and manipulated within a local scope of GP tree. Vectorization consists in grouping input variables and processing them as vectors. We demonstrate these extensions, along with an island model, allow to evolve general perceptron-like programs, i.e. working for any number of inputs. This is unlike in standard GP, where inputs are represented explicitly as scalars, so that scaling up the problem would require to evolve a new solution. Moreover, we find vectorization allows to represent programs more compactly and facilitates the evolutionary search. © 2010 Springer-Verlag.