A Machine Learning Approach for the Bin Packing Problem

Among machine learning paradigms, reinforcement learning aims to train an agent to operate in a dynamic environment in order to maximize the overall reward. By choosing in an appropriate way the reward, it is possible to find optimal solutions for many problems. This work, using a new reward concept, aims to train an agent to imitate a reference heuristic. In particular, the reward is proportional to the agent's ability to make the same choices of a particular heuristic, when applied to a given problem state. The proposed strategy is used to address the bin packing problem. The collected computational results show the validity of the proposed approach and the ability of the agent to outperform the reference algorithm.