Multi-Task Multi-Agent Reinforcement Learning With Interaction and Task Representations.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Multi-Agent Systems

Background:

• Multi-task multi-agent reinforcement learning (MT-MARL) leverages knowledge across tasks for improved performance.
• Current methods learn independent policies on shared representations, limiting explicit interaction and task relation analysis.

Purpose of the Study:

• To propose Representing Interactions and Tasks (RIT), a novel MT-MARL algorithm.
• To explicitly characterize intra-task agent interactions and inter-task task relations for effective knowledge transfer.

Main Methods:

• RIT employs interactive value decomposition to model agent dependencies, approximating Shapley values for interaction representation.
• Task representations are learned from agent trajectories to identify task similarities and relations.
• A universal policy structure supports scalable multi-task learning.

Main Results:

• RIT effectively transfers interaction knowledge across similar multi-agent tasks.
• Evaluations show significant performance improvements against state-of-the-art baselines.
• The algorithm demonstrates effectiveness in both multi-task and zero-shot learning settings.

Conclusions:

• RIT offers a novel approach to MT-MARL by explicitly modeling agent interactions and task relations.
• The method enhances knowledge transfer and scalability in cooperative multi-agent systems.
• RIT achieves superior performance, highlighting its effectiveness in complex MT-MARL scenarios.