Channon, A. D., and R. I. Damper. “Evolving novel behaviors via natural selection.” Proceedings of Artificial Life VI, Los Angeles (1998): 384-388.
URL URL2
The traditional fitness function based methodology of artificial evolution is argued to be inadequate for the construction of entities with behaviors novel to their designers. Evolutionary emergence via natural selection (without an explicit fitness function) is the way forward. This paper further considers the question of what to evolve, the focus being on principles of developmental modularity in neural networks. To develop and test the ideas, an artificial world containing autonomous organisms has been created and is described. Results show the developmental system to be well suited to long-term incremental evolution. Novel emergent strategies are identified both from an observer’s perspective and in terms of their neural mechanisms. How to Evolve Novel Behaviors The Artificial Life goal presents us with the problem that we do not understand (natural) life well enough to specify it to a machine. Therefore we must either increase our understanding of it until we can, or create a system which outperforms the specifications we can give it. The first possibility includes the traditional top-down methodology, which is clearly as inappropriate for ALife as it has proved to be for AI. It also includes manual incremental (bottom-up) construction of autonomous systems with the aim of increasing our understanding and ability to model life by building increasingly impressive systems, retaining functional validity by testing them within their destination environments. The second option is to create systems which outperform the specifications given them and which are open to producing behaviors comparable with those of (albeit simple) natural life. Evolution in nature has no (explicit) evaluation function. Through organismenvironment interactions, including interactions between similarly-capable organisms, certain behaviors fare better than others. This is how the non-random cumulative selection works without any long-term goal. It is why novel structures and behaviors emerge. As artificial evolution is applied to increasingly complex problems, the difficulty in specifying satisfactory evaluation functions is becoming apparent – see (Zaera, Cliff & Bruten 1996), for example. At the same time, the power of natural selection is being demonstrated in prototypal systems such as Tierra (Ray 1991) and PolyWorld (Yaeger 1993). Artificial selection involves the imposition of an artifice crafted for some cause external to a system beneath it while natural selection does not. Natural selection is necessary for evolutionary emergence but does not imply sustained emergence (evermore new emergent phenomena) and the question “what should we evolve?” needs to be answered with that in mind (Channon & Damper 1998). This paper sets out to answer that question. Further discussion concerning evolutionary emergence can be found in (Channon & Damper 1998), along with evaluations of other natural selection systems. Note that an explicit fitness landscape is not a requirement for artificial selection and so an implicit fitness landscape does not imply natural selection. General issues concerning long-term evolution have been addressed by Harvey’s ‘Species Adaptation Genetic Algorithm’ (SAGA) theory (Harvey 1993). He demonstrates that changes in genotype length should take place much more slowly than crossover’s fast mixing of chromosomes. The population should be nearly-converged, evolving as species. Therefore the fitness landscape (actual or implicit) must be sufficiently correlated for mutation to be possible without dispersing the species in genotype space or hindering the assimilation by crossover of beneficial mutations into the species.