Relearning and Evolution in Neural Networks
Inman R. Harvey · Adaptive Behavior · 1995
of neural networks that evolve (to get tter at one task) at the population level and may also learn (a di erent task) at the individual level. One result stated was that average tness at the evolutionary task is improved when lifetime learning at the di erent task is introduced. A di erent explanation will be proposed here for much of the data there presented: that the main results are an artefact of the unconventional evolutionary algorithm used, and can be interpreted rather di erently as a form of relearning. Asexual evolution (mutation only) was used on a population of 100 individuals or animats � the connection weights were genetically speci ed for a feedforward network for each individual, which transformed sensory inputs of the animat into movements over a grid-likeenvironment on which food had to be found. Mutation of o spring of selected parents perturbed the values of 5 of their weights chosen at random. The selective pressure used was exceptionally strong. Whereas in population genetics selective di erences are typically of the order of 1%, and with conventional genetic algorithms selective pressures are kept low toavoid premature convergence, here the ttest members have 500%moreo springthanthe average | the top 20 out of 100 each have5 o spring. In the absence of mutation such selection results in the elite taking over the whole population in just 3 generations (from 1 % to 5 % to 25 % to 100%).