LEARNING TO GROW: A LEARNING-BASED MULTI-AGENT ENDOGENOUS GROWTH MODEL

Abstract

To explain economic growth, economists typically resort to a simple and stylized optimal control problem. This problem is solved by a representative agent, who plans, at a prespecified initial date, how to allocate resources over time, between consumption and savings, with the objective of maximizing intertemporal utility. It is assumed that the agent is rational, well informed, and capable of planning over long horizons (eventually an infinite horizon). In this study, the benchmark economic growth model is revisited and reinterpreted after relaxing some of its key assumptions. In particular, the individual agent in the adapted growth model will not address a single longterm problem; instead, a sequence of shortrun decisions has to be pondered. Moreover, the recurrent decisions on how much to save, that the agent will have to make, are shaped by a learning process emerging from the systematic comparison between the agent?s own utility and the utility accomplished by other agents in the system. Additionally, it is assumed that the individual agent has imperfect knowledge about the choices of others, what introduces an element of stochasticity into the model. Simulations reveal that approaching standard growth theory through the lens of a learningbased multiagent system offers important new insights about the growth process, including an opportunity to integrate in a single framework longterm growth and shortterm irregular and unpredictable business fluctuations.