Model the Real, Artificial, or Stylized Iguana? Artificial Life and Adaptive Behavior Can Be Linked Through Pattern-Oriented Modeling

How are the disciplines adaptive behavior and artificial life related? They seem to overlap a lot, addressing adaptive behavior of organisms: how did it evolve, how does it work, what consequences does it have? Ideally, these disciplines could be linked, combining their perspectives and approaches to mutual benefit. There are, however, important barriers to this linkage. In her target article Webb identifies two main reasons for this: Artificial life (ALife) and adaptive behavior (AB) are both forms of modeling, but ALife does not explicitly follow the protocols of modeling. And, ALife denies reference to real organisms and instead focuses on artificial life to distill general principles of life, evolution, and adaptive behavior. Webb believes that these characteristics of ALife severely limit its contribution to understanding real biological phenomena. We comment on Webb’s article from the perspective of ecological modelers with a strong interest in both AB and ALife. The reason for this interest is the now widespread recognition that adaptive behavior can have strong effects on populations, communities, and ecosystems. To understand how behavior affects higher level ecology, a new type of ecology is emerging: individual-based ecology (Grimm & Railsback, 2005). A key element of individual-based ecology is the use of individual-based models in which populationand community-level ecology emerge from individual-level adaptive behavior. Some examples are models of how the effects of habitat alteration on fish populations are mediated by individuals selecting habitat as a trade-off between predation risk and growth (Railsback & Harvey, 2002); and of how population persistence of marmots is affected by individual dispersal decisions that affect survival and probability of reproduction (Grimm et al., 2003). Webb’s first point is that ALife does in fact use models—be they robots or computer agents—but does not explicitly use a modeling research program. To evaluate this point, we need to clearly define what we mean by model. A general definition is given by Starfield, Smith, and Bleloch (1990; Grimm & Railsback, 2005): A model is a purposeful representation. A model needs to have a purpose because otherwise there would be no way to decide what to include in it. A model’s purpose is a filter: the model should not include anything not believed essential for explaining the phenomenon of interest (Starfield et al., 1990). Then, in an iterative modeling cycle (Grimm & Railsback, 2005) the model is implemented, tested, compared with the phenomenon of interest, then revised, re-tested, and so on, until both the phenomenon of interest is sufficiently reproduced and the mechanisms generating the phenomenon are sufficiently understood. In science, the problem addressed by modeling is usually to understand some specific phenomenon. Webb calls this the target phenomenon and convincingly makes the case that ALife certainly wants to