MO-MIX: Multi-Objective Multi-Agent Cooperative Decision-Making With Deep Reinforcement Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Robotics

Background:

• Deep reinforcement learning (RL) excels at complex decision-making.
• Real-world tasks often involve multiple conflicting objectives and cooperative agents.
• Existing research inadequately addresses multi-objective multi-agent reinforcement learning (MOMARL).

Purpose of the Study:

• To propose MO-MIX, a novel framework for solving MOMARL problems.
• To enable cooperative decision-making in scenarios with multiple conflicting objectives.
• To generate an approximation of the Pareto set for MOMARL tasks.

Main Methods:

• Utilizes the centralized training with decentralized execution (CTDE) framework.
• Incorporates a weight vector to condition local action-value function estimation.
• Employs a parallel mixing network for joint action-value function estimation.
• Applies an exploration guide for improved non-dominated solution uniformity.

Main Results:

• MO-MIX effectively solves multi-objective multi-agent cooperative decision-making problems.
• The method generates a high-quality approximation of the Pareto set.
• Demonstrates significant performance improvements over baseline methods across four evaluation metrics.
• Achieves superior results with lower computational cost.

Conclusions:

• MO-MIX provides an effective solution for MOMARL problems.
• The proposed approach advances the state-of-the-art in cooperative AI.
• Offers a computationally efficient and high-performing method for complex decision-making.