Affiliations: School of Electrical and Computer Engineering, Shiraz University, Shiraz, Iran | Iranian Space Research Center, Institute of Mechanics, Shiraz, Iran
Note: [] Corresponding author. Behrooz Raeisy, E-mails: raeisy@shirazu.ac.ir; golbahar@shirazu.ac.ir (Shapoor Golbahar Haghighi); safavi@shirazu.ac.ir (Ali Akbar Safavi).
Abstract: Multi-Agent systems have proved powerful in various sciences and engineering problems. This paper proposes a novel Multi-Agent Active Noise Control (ANC) formulation via the credit assignment approach. The introduced ANC removes multi-tonal acoustic noises in the environment invoking reinforcement learning techniques. In some multi-agent systems, for the training of all agents, only one reward is available. It is clear that this reward does not belong to one particular agent. The assignment of this reward to the agents is a problem which is known as Multi-Agent Credit Assignment (MCA). In this research, each agent is responsible for reducing the noise power of one single harmonic, while only the total noise power of the signals is known. Therefore, it is required to assign a power contribution to each single harmonic. To resolve this problem, at first, the Knowledge Evaluation Based Critic Assignment (KEBCA) idea with proper modification is used and then a new method is introduced for this special problem. Simulation results show good improvement in the system performance by switching the single agent into the multi-agent system.
Keywords: Active Noise Control (ANC), Knowledge Evaluation Based Critic Assignment (KEBCA), Multi-Agent Credit Assignment (MCA), Multi-agent system, Q-Learning (QL), Reinforcement Learning (RL)
