Lessons from ethology for computational models of development

Summary form only given. Recent results from the ethological literature challenge some widely held views about the development of animal behavior and offer new directions for research into the organization of synthetic behavior systems. The process of development is often seen as one of continual refinement, with infantile behaviors representing the primitive precursors of adult behavior. While this notion is intuitively pleasing, there are two ways in which it may be incorrect. First, Coppinger (Coppinger and Smith 1990) argues that infantile behaviors are just as well adapted as adult behaviors, but adapted to the challenges that the organism faces during its early life and early social context. In his view, development is best understood as a process of, 'swapping' elements in the behavioral repertoire rather than refining them; the adult version of a behavior, in other words, is a distinct replacement for its juvenile counterpart, not an outgrowth of it. The second problem with the traditional view of behavioral development is that infantile behaviors are often expressed in an entirely different motivational context than their adult counterparts. Indeed, infantile behaviors that satisfy a particular motivational system such as hunger may be controlled by entirely different motivational systems. Hall and Willams have shown, for example, that during the first several weeks of life, a rat pup's suckling rate is independent of how hungry it is. That is, rat pups suckle because they are innately motivated to do so, not because they are hungry. Similarly, Hogan (Hogan 1999) suggests that during the first 48 hours of a chick's life, the amount that it pecks has nothing to do with its level of hunger. Leyhausen (Lorenz and Leyhausen 1973) has shown that the behaviors that make up the sequence of predatory behaviors in cats are refined through play, often well before the cat ever makes 'the connection' between the acquisition of prey and a subsequent reduction in hunger. What is going on here? In the first two cases, the juvenile behaviors act as 'fail-safe' mechanisms to ensure that the creature's needs are met without needing an innate or learned connection between the behaviors and the critical motivational systems that they act to satisfy. In the latter case, behaviors that will be critically important in adulthood are initially expressed and refined in a context in which the costs of failure are low, i.e. the creature's life doesn't depend on how well they perform the behavior at that given moment. Both of these concerns suggest that by re-examining our assumptions about development in animals, we may be able to glean generally useful organizing principles for the design of adaptive computational systems.