Animats in the Modeling Ecosystem

There are many different kinds of model and scientists do all kind of things with them. This diversity of model type and model use is a good thing for science. Indeed, it is crucial especially for the biological and cognitive sciences, which have to solve many different problems at many different scales, ranging from the most concrete of the structural details of a DNA molecule to the most abstract and generic principles of self-organization in networks. Getting a grip (or more likely many separate grips) on this range of topics calls for a teeming forest of techniques, including many different modeling techniques. Barbara Webb’s target article strikes us as a proposal for clear-cutting the forest. We think clear-cutting here would be as good for science as it is for non-metaphorical forests. Our argument for this is primarily a recitation of a few of the ways that diversity has been useful. Recently, looking at the actual practice of artificial life modelers, one of us distinguished four uses of simulation models classified in terms of the position the models take up between theory and data (see Figure 1). The classification is not exhaustive, and the barriers between kinds are not absolute. Rather, the purpose of the taxonomy is to open up the view for an epistemic ecology of modeling practices. First, and closest to the empirical domain, there are mechanistic models, in which there is an almost one-to-one correspondence between variables in the model and observables in the target system and its environment. Webb’s cricket robot is a paradigmatic example of this type. Second, there are functional models, which aim for a behavioral or functional rather than a variable-to-variable correspondence between the model system and its target. Many models in cognitive psychology are of this type, as are many in biology when the underlying mechanisms are not accessible to modeling correspondence (see Vickerstaff & Di Paolo, 2005, for a good example of this). Third, there are generic models, which cover a wide spectrum of phenomena in search for generic principles of complex systems. Cellular automata, random Boolean networks, and the like belong to this class. Finally, there are conceptual models, which do not target any particular natural system nor a wide spectrum of them. Instead conceptual models are built from theories, for which they embody assumptions, illustrate concepts, simulate theoretical principles, and so forth. Beer’s model and many other animats are of this type (for more details see Barandiaran, 2008, chap. 2; Barandiaran & Moreno, 2006). Modeling is a relational activity, it involves the template or construct (usually referred as the model) and also an interpretative framework (made of assumptions, generalizations, definitions, etc.) that puts the model in connection with other models, theories or objects. It is this interpretative framework and the modelers’ intentions that situate a given model into one of the categories of models. But only mechanistic models are used as Webb suggests all models should be: “the cricket robot is a mixed physical and computational implementation of a particular hypothesis about an observed phenomenon of sensorimotor behavior in