Skip to search boxSkip to navigationSkip to main content

On the Benefits of Divergent Search for Evolved Representations

Research Output:
Conference Article in Proceeding or Book/Report chapter
Article in proceedings
Research
Peer-review

Abstract

Evolved representations in evolutionary computation<br/>are often fragile, which can impede representation-dependent<br/>mechanisms such as self-adaptation. In contrast, evolved representations<br/>in nature are robust, evolvable, and creatively exploit<br/>available representational features. This paper provides evidence<br/>that this disparity may partially result from a key difference<br/>between natural evolution and most evolutionary algorithms:<br/>Natural evolution has no overarching objective. That is, nature<br/>tends to continually accumulate novel forms without any final<br/>goal, while most evolutionary algorithms eventually converge to<br/>a point in the search space that locally maximizes the fitness<br/>function. The problem is that individuals that maximize fitness<br/>do not need good representations because a representation’s<br/>future potential is not reflected by its current fitness. In contrast,<br/>search methods without explicit objectives that are consequently<br/>divergent may implicitly reward lineages that continually diverge,<br/>thereby indirectly selecting for evolvable representations that<br/>are better able to diverge further. This paper reviews a range<br/>of past results that support such a hypothesis from a method<br/>called novelty search, which explicitly rewards novelty, i.e.<br/>behaviors that diverge from previously encountered behaviors.<br/>In many experiments, novelty search demonstrates significant<br/>representational advantages over traditional fitness-based search,<br/>such as evolving more compact solutions, uncovering more evolvable<br/>representations, and more fully exploiting representational<br/>features. The conclusion is that divergent evolutionary algorithms<br/>like novelty search may exert selection pressure towards higher<br/>quality representations than traditional convergent approaches<br/>to search.

Publication Information

Output type

Research Output:
Conference Article in Proceeding or Book/Report chapter
Article in proceedings
Research
Peer-review

Original language

English

Publication milestones

  • Published - 2012

Publication status

Published - 2012

Book series

  • Book series name: Proceedings of the EvoNet 2012 Workshop at ALIFE XIII

Host publication title

Proceedings of the EvoNet2012 Workshop at the Thirteenth International Conference on Artificial Life (ALIFE XIII)