Mean-Variance Preferences as the Outcome of Reinforcement Learning

I show that common assumptions about learning leads to risk averse be- havior: in the long run the learner is most likely to choose the least variable alternative among a set of alternatives with identical expected values. If the learner can choose between normally distributed alternatives, reinforcement learning leads to behavior identical to that of risk averse decision makers who have mean-variance preferences and choose between the alternatives according to a logit choice rule. A modied learning rule generates behavior consistent with an s-shaped utility function.