The Baldwin Effect

Finally, the long-awaited topic to breach the publish button! Read a great editorial: Evolution, Learning, and Instinct: 100 Years of the Baldwin Effect. In 1896, with Lamarckism being discredited, James Mark Baldwin put forth his own evolutionary corollary. He said, first, that there exist evolutionary pressures towards phenotypic plasticity, but second, that there are also pressures away from plasticity, and towards rigid hard-wiring.

The Baldwin Effect has particular bearing on evolutionary computation, and in fact seems to seldom appear in other contexts. It's probably passed over as an incidental facet of genetic assimilation (which it is) but we benefit by singling it out as an independent phenomenon. The typical, trivial examples are tanning and the development of calluses; for me and for most, however, it's a statement on learning.

Within the algorithmic exploration of a fitness landscape, the Baldwin Effect separates local search from global search. Genetic disposition drops you somewhere, and there you are. Nothing you can do about it. But in a dynamic environment, on bumpy fitness terrain, there's great benefit to developing learning. It means agents can adaptively (and stochastically) learn and explore the local fitness landscape.

But it's no cakewalk! That's Baldwin's second point. Sometimes learning is the wrong approach, perhaps because it's too risky to experiment, or too wasteful, or too flighty. Turney (1996) has a great cost/benefit analysis of learning versus instinct.

A challenge: natural selection coupled with learning is very worthwhile, but better is the emergence of learning through natural selection. We want the system that adapts the plasticity of its plasticity depending on fitness landscape; we want to see learning as subject to genetics.