Evolution of Risk Aversion in Adaptive Learning Agents

Risk aversion is one of the most commonly cited properties of human decision making (e.g., Kahneman & Tversky, 1979). This finding is at odds with traditional expected value theory, but not with more recent theories of rational choice (e.g., von Neumann & Morgenstern, 1944). Since Bernoulli?s solution of the St. Petersburg paradox by appeal to marginally decreasing sensitivity to money (his so-called ?moral expectation?), there has been a debate as to the rationality of this apparent risk aversion. Rather than evaluating rationality by appealing to notions of consistency (as is often done), another, I think, more satisfying approach is to evaluate the survival/adaptive advantage conferred on an individual who exhibits particular behaviors: If one finds that a behavior x gives a greater survival advantage than behavior y, then, one can reasonably argue, the latter is more rational than the former. To investigate the adaptive advantage of different risk preferences, I used a replicator dynamic to model the evolution of populations of adaptive learning agents (Roth-Erev machines) in repeated coordination games. Risk aversion consistently displaced risk seeking and risk neutral behavior in the populations, and, as a result, the populations evolved to prefer the risk dominant (inefficient) equilibrium point. These results support the argument that risk aversion (at least for gains) is a rational property of adaptive learning agents (and, by extension, of people).