Learning to be risk averse?

The purpose of this research is to search for the best (highest performing) risk profile of agents who successively choose among risky prospects. An agent's risk profile is his attitude to perceived risk, which can vary from risk preferring to risk neutral (an expected-value decision maker) to risk averse. We use the Genetic Algorithm to search in the complex stochastic space of repeated lotteries. We find that agents with a CARA utility function learn to possess risk-neutral risk profiles. Since CARA utility functions are wealth-independent, this is not surprising. When agents have wealth-dependent, CRRA utility functions, however, they also learn to possess risk profiles that are about risk neutral (from slightly risk-averse to even slightly risk-preferring), which is surprising.