Learning by Genetic Algorithms in Economic Environments

This paper studies models in which economic agents use genetic algorithms to learn about their environment and to learn about their objective functions. The main results of the paper, obtained through the analysis of computer simulations, are: For a model in which competitive firms have to make their production decisions before observing the price of a single good in a market, the values of prices and quantities to which the algorithm converged correspond to rational expectations equilibria. The beliefs of all firms about how much to produce and offer for sale converge to the same value which is equal to the optimal quantities if the market price is known. For an overlapping generations model with a constant stock of money (in which agents make decisions about how much to consume of a single, non-storable good in the first period of their life, without knowledge of the price of a good in a second period), the algorithm converges to the unique monetary equilibrium (the moving average learning scheme converges to the same equilibrium and the results from experimental overlapping generations environments are in the domain of attraction of this equilibrium). The agents form beliefs about the consumption in the first period as if they learned to maximize their utility functions and learned the correct price prediction. For an overlapping generations model with constant deficit financed through seignorage (where agents make the same decision as in the previous model), the algorithm converges to the low-inflation-rate equilibrium (which is stable under least-squares learning scheme and which is the domain of attraction for the convergene of experimental economies). The genetic algorithm also converges to this equilibrium for the values of deficit and initial inflation rate for which the least-squares scheme diverges. In a simple asset market model where a group of traders learn about the relationship between the price and the return on asset, using genetic algorithms to determine their estimates of the return, price and quantities traded converge to rational expectations price and quantities. Convergence of genetic algorithm to rational expectations equilibria was also obtained for the values of a stability parameter for which least squares do not converge.