Learning to deal with risk: what does reinforcement learning tell us about risk attitudes?

People are generally reluctant to accept risk. In particular, people overvalue sure gains, relative to outcomes which are merely probable. At the same time, people are also more willing to accept bets when payoffs involve losses rather than gains. I consider how far adaptive learning can go in explaining these phenomena. I report simulations in which adaptive learners of the kind studied in Roth & Erev (1995, 1998) and Borgers & Sarin (1997, 2000) deal with a problem of repeated choice under risk where alternatives differ by a mean preserving spread. The simulations show that adaptive learning induces (on average) risk averse choices. This learning bias is stronger for gains than for losses. Also, risk averse choices are much more likely when one of the alternatives is a certain prospect. The implications of a learning interpretation of risk taking are explored.