Evolution and Learning in Neural Networks: Dynamic Correlation, Relearning and Thresholding
Brian Carse, Johan Oreland · Adaptive Behavior · 2000
This contribution revisits an earlier discovered observation that the average performance of a pop ulation of neural networks that are evolved to solve one task is improved by lifetime learning on a different task. Two extant, and very different, explanations of this phenomenon are examined- dynamic correlation, and relearning. Experimental results are presented which suggest that neither of these hypotheses can fully explain the phenomenon. A new explanation of the effect is proposed and empirically justified. This explanation is based on the fact that in these, and many other relat ed studies, real-valued neural network outputs are thresholded to provide discrete actions. The effect of such thresholding produces a particular type of fitness landscape in which lifetime learn ing can reduce the deleterious effects of mutation, and therefore increase mean population fitness.