Extrapolation of regularity using indirect encodings

Abstract

The choice of training data used in evolution can have a significant impact on the generalized performance of the evolved solutions. Historically, if the training set was not representative of the problem's overall state space, the evolved solutions could not practically be applied to the overall problem. However, generative systems and indirect encodings are able to identify and leverage regularities in and the geometry of the state space to produce effective solutions to complex problems. This ability presents the possibility of using the regularity of a problem to effectively extrapolate evolved solutions to areas of the state space for which the training set was not representative. In this work, two different experiments are performed involving pattern reproduction and robot control to explicitly evaluate this extrapolation ability. Results show that an indirect encoding is able to extrapolate performance in one area of a problem's state space to a new area in which it has no experience with little to no loss of performance, depending on the regularities of the problem's state space. If the regularities were consistent through the entire state space and across the boundary between areas in which there was experience and no experience, extrapolation performance was high, but if the regularities were not consistent, there was a loss of performance.