An Entanglement-Inspired Action Selection and Knowledge Sharing Scheme for Cooperative Multi-agent Q-Learning Algorithm used in Robot Navigation
Mohammad Hasan Karami, Hossein Aghababa, Amir Hosein Keyhanipour · 2020
Multi-agent reinforcement learning, especially learning in unknown complex environments, requires new algorithms. In this work, our focus is on adopting the concept of the quantum entanglement phenomena to the action selection procedure of multi-agent Q-learning, aiming to enhance the learning speed, collision avoidance, and also providing full coverage of the environment. The exploration procedure is exclusively induced by a memory-based probabilistic sequential action selection method acting as a knowledge hub, shared among the agents, which is the central pillar of this work. This causes enhancing the parallelism of the learning process, plus, building an effective yet simple communicating-bridge between the learning agents; that is, they can signal and guide one another through sharing their gained experience and knowledge in order not to repeat the same mistake that the other agents have already run into. The simulation results demonstrated the effectiveness of our proposed algorithm in terms of reducing the learning time, significant reduction of collision occurrence, and providing full coverage of big complex clutter environments.