Towards the evolution of scaleable neural architectures

Abstract

There has been a great deal of interesting work published on evolving neural networks in the last few years, some of which is mentioned below. Nearly all previous work, however, has concentrated on evolving a particular neural network to solve a particular problem. When a suitable solution is evolved, then all we have is a suitable solution for a particular problem. The work reported here offers a significant departure from that theme, and presents a simple system which allows the evolution of scaleable neural architectures. This is important for two reasons: Evolutionary search is computationally expensive. When evolving solutions to complex problems, it might be better to evolve solutions to small examples of the problem, then for the real application, scale up some of the best evolved solutions to the real problem size. Having evolved good solutions to a problem, it would be good to apply them to similar problems of a different size. By allowing this, we allow maximum possible reuse of neural network modules.