BEAUT: An ExplainaBle Deep LEarning Model for Agent-Based PopUlations With Poor DaTa

Abstract

When working with an agent-based system, it may be desirable to develop an algorithm that can predict future time periods for each individual. One such approach that works to accomplish this is the Markov decision process, which takes as its input available actions, unique agent reward functions, and a model of the overall environment. However, it can be difficult to identify the underlying factors that influence decision making when attempting to simulate the behavior of real-world populations. For example, when modeling the transitions of homeless individuals between states such as street and shelter, it can be a challenging task as the external factors impacting them (e.g., weather) may not be readily apparent. Therefore, this paper proposes and evaluates an approach to capture this information in an explainable way from aggregate, real-life data produced by the At Home/Chez Soi project. The proposed algorithm “BEAUT” is a consolidation of a modified deep q-learning (MDQL) and modified neural fitted q-iteration (MNFQ) algorithm that work together to generate a set of probabilistic transition matrices to describe state transitions. BEAUT is evaluated with experimental results that compare its accuracy against similar methods for time-series forecasting using real-world data. Our tests show that BEAUT provides an explainable forecasting model without a loss of accuracy, and in some instances results in higher accuracy.