Research map / Multi-agent systems and games / Cooperative multi-agent RL
Explanations for Multi-Agent Reinforcement Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Doctoral consortium proposal outlining post-hoc explanation methods for centralized and decentralized multi-agent reinforcement learning (MARL), including policy summaries, query-based and temporal explanations.
Problem
Most explainable RL work targets single-agent settings, leaving post-hoc explanations for centralized and decentralized MARL policies underdeveloped.
Approach
Build abstract MMDP representations of centralized MARL policies to produce policy summaries, query-based explanations (via minimal Boolean logic) and temporal explanations (via PCTL* model checking); for decentralized MARL, aggregate agents' behaviors with Hasse diagrams and present them via position overlays in AR, with planned LLM-based condensation of explanations.
Key results
- Centralized methods reportedly scale to up to 19 agents for summaries and queries and up to 9 agents for temporal explanations, with user studies showing improved performance and goodness ratings.
- Decentralized position-overlay AR presentation reportedly increased user performance and satisfaction in a user study supported by another student's work.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the centralized policy summarization and query-based explanation methods | Centralized MARL · Three MARL domains; real-world user study | user performance | Helped |
| Use the centralized policy summarization and query-based explanation methods | Centralized MARL · Real-world user study | subjective explanation goodness | Helped |
| Use the temporal explanation method | Centralized MARL · Four domains | user performance | Helped |
| Use the temporal explanation method | Centralized MARL · Four domains | subjective goodness | Helped |
| Use decentralized policy summarization with position overlay | Decentralized MARL · AR-based user study | user performance | Helped |
| Use decentralized policy summarization with position overlay | Decentralized MARL · AR-based user study | user satisfaction | Helped |
Limitations
- Decentralized query-based explanations are still under development and may be lengthy with redundant information.
- Planned computational experiments and user studies are not yet complete.
Experimental setup
Builds on: Topin and Veloso 2019 (abstract MDP policy summarization) Hayes and Shah 2017 (query-based explanations) Sreedharan et al. 2022 (temporal explanations)
Closest papers by meaning
- UTILITY: Utilizing Explainable Reinforcement Learning to Improve Reinforcement Learning (ICLR 2025 · similarity 0.89)
- Explaining Reinforcement Learning Agents through Counterfactual Action Outcomes (AAAI 2024 · similarity 0.87)
- CrystalBox: Future-Based Explanations for Input-Driven Deep RL Systems (AAAI 2024 · similarity 0.81)
- Approximating Shapley Explanations in Reinforcement Learning (NeurIPS 2025 · similarity 0.81)
- A Principle of Targeted Intervention for Multi-Agent Reinforcement Learning (NeurIPS 2025 · similarity 0.80)
- Explaining Decisions of Agents in Mixed-Motive Games (AAAI 2025 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.