The baldwin effect hinders self-adaptation

Abstract

The “end-game” of evolutionary optimisation is often largely governed by the efficiency and effectiveness of searching regions of space known to contain high quality solutions. In a traditional EA this role is done via mutation, which creates a tension with its other different role of maintaining diversity. One approach to improving the efficiency of this phase is self-adaptation of the mutation rates. This leaves the fitness landscape unchanged, but adapts the shape of the probability distribution function governing the generation of new solutions. A different approach is the incorporation of local search – so-called Memetic Algorithms. Depending on the paradigm, this approach either changes the fitness landscape (Baldwinian learning) or causes a mapping to a reduced subset of the previous fitness landscape (Lamarkian learning). This paper explores the interaction between these two mechanisms. Initial results suggest that the reduction in landscape gradients brought about by the Baldwin effect can reduce the effectiveness of self-adaptation. In contrast Lamarkian learning appears to enhance the process of self-adaptation, with very different behaviours seen on different problems.