Evolution, Learning, and Culture: Computational Metaphors for Adaptive Algorithms

Potential interactions between connectionist learning syst ems and algorithms modeled aft er evolutionary adapt ation are becoming of increasing interest . In a recent short and elegant paper Hinton and Nowlan extend a version of Holland's genetic algorithm (GA) t o consider ways in which the evolution of species and th e learning of individu als might in teract [17]. Their mod el is valuable both becaus e it provides insight into potenti al interactions between th e natu ral processes of evolution and learning and as a potential bridge between the arti ficial questions of efficient and effective machine learning using t he GA and connectioni st networks. Thi s pa per begins by describing the GA and Hinto n and Nowlan's simulati on . We then an alyze their model, use thi s analysis to explai n its nontrivial dynamical behavio rs, and consider the sensitivity of the simulation to several key parameters. Our next step is t o in terpose a thi rd adapt ive system culture between the learning of individuals and the evolut ion of pop ulations . Culture accumulates the "wisdom" of individuals' learning beyond the lifetime of any one indi vidual but adapts more responsively than the pace of evolut ion allows. We describe a series of exper iments in which the most minimal notion of culture has been ad ded to the Hinton and Nowlan mod el, and we use, this experience to comment on the functional value of cult ure, and similarities between an d interac tions among these three classes of adaptive systems.