Opening a New Frontier in Social Simulation with Machine Learning

Abstract

In this paper, we introduce an application of machine learning to simulation optimization, in which the simulation is used as an evaluator to search for optimal solutions. This is one of the efforts to integrate social simulation with machine learning. To find the optimal solutions using simulations as an evaluator, the execution of thousands or tens of thousands of simulations is required. Therefore, it is an important issue to reduce the computational cost. To reduce the computational cost, using a lightweight surrogate model replicating the simulation is useful. We have developed a surrogate model building framework for complex social simulations, like multi-agent simulations, by using deep neural networks. We have applied the proposed method to an urban-scale traffic simulation and demonstrated that it can simulate each road density as well as the traffic simulation.