The Learning Power of Evolution
Vitaly Feldman, Leslie Gabriel Valiant · 2009
It has been widely recognized that learning and evolution have the commonality of involving adaptive processes that once started do not need a programmer or designer. It is tempting to seek some mystical extra power in evolution, beyond that of learning, simply because of the apparently spectacular consequences of evolution that we see around us. However, such approaches have not succeeded to date. In response to this situation one of the authors made the, apparently radical, suggestion that evolution is nothing other than a constrained form of computational learning. In [Val08] a notion of evolvability was defined in a similar spirit to the definition of learnability. The goal of the definition is to offer a rigorous basis for the analysis of evolution and for distinguishing between efficient evolution and evolution that is only realized in some exponentially far limit.