A Fairness-Aware Cooperation Strategy for Multi-Agent Systems Driven by Deep Reinforcement Learning

The research on multi-agent cooperation strategies has been attracting widespread concerns in recent years. However, the current deep reinforcement learning algorithms mainly focus on improving cooperation efficiency while ignore fairness. Taking into account both collaboration efficiency and fairness is a complex multi-objective optimization problem. To address this concern, we design a Fair-Efficiency Multi-Agent Deep Deterministic Policy Gradient (FE-MADDPG) algonthm. First, we design a fair and efficient reward function which sets the resource occupancy rate as the ratio of each agent's average reward to the total reward to ensure the fairness of each agent. Then, we improve the MADDPG algonthm by utlhzmg the reward function and make comparisons of the efficiency of agents. Finally, we employ the Gini coefficient and the time consumed for completing the task as evaluation indicators to verify the fairness and efficiency. Simulation results show that the FE-MADDPG algonthm significantly improves the efficiency of the system under the premise of ensuring fairness for each agent.