walter wrote:
One of the easiest problems and very interesting problems is the study of the interaction between 'learning' and evolution, and in particular the baldwin effect. This problem is relatively 'easy' from a implementationary perspective, since we model 'learning' simply as a plasticity of a trait (in this case rotation and forward speed) during life - not necessarily in a positive way.
This is inspired by Hinton & Newlon (1990?) who hold that learning/plasticity can 'guide' evolution by smoothing the fitness landscape (a spike is smoothed to a bell) and thereby better facilitating easier evolutionary search. Others have proposed that this effect is particularly useful in changing fitness landscape (no spike, but moving optima). The baldwin effect has thus taken place if phenotypic plasticity has facilitated a genotypical change.
In our experiments, we combine the coevolutionary dynamics between predator and prey with a study to this effect. The coupled fitnesses of the two species (coevolution) takes care of a dynamically changing fitness landscape. Individuals get a gene that switched the plasticity a rotation or forward speed on or off, which models learning.
The genotype of an individual in simple coevolution consists of two floating points (neural properties that determine the rotation and forward speed). The genotype of individuals in learning/coevolution experiments have two additional 'boolean' genes (plasticity: on/off, direction of plasticity: faster/slower).
The premilinary results of such experiments are very promising :-) By comparing the population dynamics of the ecosystem without mutation (all clones) with the population dynamics of a coevolving ecosystem, we can easily (graphically) spot changes in the equilibrium/attractor that are due to genotypic change. By analysing the genotypes and life histories of the evolving populations we can identify the genotypical mutations that have lead to these changes in population dynamics. Moreover, the adaptations (changes in speeds) of the two species show a coupling that reminds to the red queen effect (or an arms race).
In the case of learning and evolution, the results seem to reflect Baldwin's theory. That is, although still many mutations are plain mutations (without guidance from evolution), some of them seem to be facilitated by phenotypic plasticity. In these cases, we can see that learning precedes this mutation (that is, the plasticity of this trait was turned on before the mutation in the same direction). And in some cases the plastic abilities are switched off afterwards. The latter makes sense if the random mutation was a good choise (near optimal) and the coupled fitness landscape is undisturbed for a while (no significant adaptative change in other species). That is, when the cost of learning (drift from optimality) exceeds its benefit, which is only a potential benefit if situated in a changing environment.