Improving evolvability through novelty search and self-adaptation

摘要: A challenge for current evolutionary algorithms is to yield highly evolvable representations like those in nature. Such evolvability natural evolution encouraged through selection: Lineages better at molding new niches are less susceptible extinction. Similar selection pressure not generally present algorithms; however, the first hypothesis this paper that novelty search, a recent technique, also selects because it rewards lineages able continually radiate behaviors. Results experiments maze-navigation domain support search finds more than regular fitness-based search. However, though outperforms second biped locomotion experiment, proves no delicately balanced behaviors fragile domain. The such fragility can be mitigated self-adaption, whereby genomes influence their own reproduction. Further domains with and self-adaption indeed demonstrate increased evolvability, while, interestingly, adding self-adaptation decreases evolvability. Thus, selecting may often facilitate when overly fragile; furthermore, achieving potential of critically depend upon reward scheme driving evolution.