Applications of evolutionary learning in macroeconomic models

Genetic Algorithms are the best known representation of a class of direct random search methods called evolutionary algorithms which are widely used to solve complex optimization and adaptation problems. They have grown in popularity within economics due to their ability to represent the adaptation of individuals to the underlying parameters of their economic system. This work examines three applications of genetic algorithm adaptation in macroeconomic environments. In the first of these applications, the Arifovic and Masson (2003) model of currency crisis is simulated in controlled laboratory experiments with human subjects. An extended model of agents expectations is considered in which each investor has multiple rules, choosing one of them probabilistically in each period. The properties of time series generated by computer simulations are compared to those of human data. In each framework the time series of returns on emerging market debt is characterized by fat tails which matches features of empirical data. Additionally, the extended model of expectations better matches the duration statistics found in the experimental setting. The second application investigates the sufficiency of learning-by-doing for explaining negative macroeconomic output shocks in an evolutionary model of technological transition. The model allows firms to divide labour between two distinct technologies in a continuous manner. The ability of each firm to innovate within each technology is dependent on this choice for the division of labour. Contrary to previous literature, innovations are not transferable between technologies. It is argued that in such a framework learning-by-doing remains sufficient for periodic observations of negative macroeconomic growth. Thc final exa~rii~lation represents the first application of two-level learning in an economic environment in which the performance of potential rules is complementary across individuals. Two-level learning, or self-adaptation, incorporates certain strategy parameters into the representation of each individual. In this work, these strategy parameters determine the level of heterogeneity introduced into the environment. They evolve by means of mutation and recombination, just as the object variables do. It is argued that self-adaptation over these parameters can replace the election operator proposed by Arifovic (1994) in order to attain convergence to a rational expectations equilibrium.