Optimizing Food Allocation in Food Banks with Multi-agent Deep Reinforcement Learning

Abstract

Food loss and waste (FLW) is becoming a more serious problem, especially in developed nations. Meanwhile, vegetables with imperfect shapes and/or appearances are often discarded even when their flavor and safety have not been compromised. This study proposes a food allocation method with a multi-agent deep reinforcement learning algorithm (called QMIX) to optimize the food allocation process in food support organizations. Several experiments are conducted in a virtual environment to evaluate the basic characteristics of the proposed method and to understand how agents can behave effectively and cooperatively (i.e., to ensure food supply works fairly) through this method. Empirical results showed that agents using the proposed method can acquire reasonable behaviors that can be applied to simple scenarios in the real world.