DQN for Coordinating Multi-agent Cooking
Yiwei Zhang · Highlights in Science Engineering and Technology · 2023
Reinforcement learning (RL) is a very widely used field. The difference between RL and other branches of machine learning (such as supervised learning and unsupervised learning) is that RL centers on interactive learning. The RL model (also known as the agent) learns in interaction with the environment to maximize the reward function. In the paper "too many cooks," the authors developed a method called Bayesian Delegation to enable human-like coordination by inferring the sub-tasks of others quickly. However, limitations still exist in the partial order of sub-tasks. First, implementing the sub-task in terms of efficient actions or which agent(s) should work on it is not specified. Second, the sub-tasks may be finished in many different orders since the ordering of sub-tasks is partial. The project proposes solutions to these challenges using Deep Q-Learning (DQN) and Bayesian Inference. In the DQN experiment, value approximation performs well in the simple multi-agent environment.