Natural Action Selection, Modeling

Put simply, action selection is the task of deciding what to do next. As a general problem facing all autonomous entities—whether animals or artificial agents—action selection exercises both the sciences concerned with understanding the biological bases of behavior (e.g., ethology, neurobiology, psychology) [Page 557]and those concerned with building artifacts (e.g., artificial intelligence, artificial life, and robotics). The problem has two parts: What constitutes an action, and how are actions selected?

Models of natural action selection allow us to test the coherence of proposed social and biological theories. Although models cannot generate data about nature, they can generate data about theories. Complex theories can therefore be tested by comparing the outcome of simulation models against other theories in their ability to account for data drawn from nature. Each model attempts to account for transitions among different behavioral options. A wide range of modeling methodologies is currently in use. Formal mathematical models have been complemented with larger scale simulations that allow the investigation of systems for which analytical solutions are intractable or unknown. These include models of artificial animals (simulated agents or robots) embedded in simulated worlds, as well as models of underlying neural control systems (computational neuroscience and connectionist approaches). A potential pitfall of more detailed models is that they may trade biological fidelity for comprehensibility.

General challenges facing models of action selection include the following: Is the model sufficiently constrained by biological data that it captures interesting properties of the target natural system? Do manipulations of the model result in similar outcomes to those seen in nature? Does the model make predictions? Is there a simpler model that accounts for the data equally well? Or is the model too abstract? Are its connections to data trivial, making it too obvious to be useful?

Models of natural action selection have delivered new insights in many domains. What follows is a review of several: (a) the relationship between evolved behavior and optimality, (b) biological mechanisms of action selection, (c) whether or not sequencing behavior can require special representations, (d) the role of perception, (e) explanations of disability or disease, and (f) finally individual action selection in a social context.