Deep learning in agent-based models

Agent-based models (ABMs) are a proven excellent computational simulation paradigm to describe complex socio-economic systems, constituted by artificial entities interacting over time within customized environments. In fact, market economies are constituted by a large number of interacting agents involved in recurrent transactions, resulting in a dynamic system whose state is ultimately defined by the uncoordinated actions of operators, based on personal expectations. ABMs can provide a good description of such a system via simulation. When the exact equations of an economic model are unknown, we might still know the behavior of the artificial agents, based on the observation of their real-world counterparts, which may be consumers, businesses, financial institutions, as well as government agencies. A computational model constructed over specific requirements can serve as a representation of a whole national economy, reproducing the development of economic indicators, such as GDP growth, inflation and unemployment rates, through the interactions of its agents. Thus, it may become a powerful instrument for policy analysis. One limitation arising from the use of AMBs is the computational cost of the simulations. This can be solved by implementing meta-models through machine learning algorithms, trained on the results of various simulations. These serve as surrogates of the original model, by reproducing the same output without the need for running the simulation. As part of our study, we developed a set of machine learning meta-models aiming at reproducing the output of the Keynes meeting Schumpeter ABM. Results indicate that deep learning methods are the most effective technique for replicating the output of the ABM.