Research map / Multi-agent systems and games / Games and equilibria
Counterfactual Effect Decomposition in Multi-Agent Sequential Decision Making
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.
Decomposes the total counterfactual effect of an agent's action in multi-agent MDPs into agent-specific effects (attributed to agents via Shapley value) and reverse state-specific effects (attributed to states via intrinsic causal contributions).
Problem
Explaining the total counterfactual effect of an agent's action on a realized trajectory's outcome in multi-agent sequential decision making, where the effect propagates both through other agents' subsequent actions and through environment state transitions.
Approach
Decomposes TCFE into a total agent-specific effect minus a reverse state-specific effect; attributes the agent-specific part to agents with Shapley values and the state-specific part to state variables via a variance-based intrinsic causal contribution measure computed with structure-preserving (noise-conditioning) interventions.
Key results
- In the Gridworld example, ASE-SV assigns the full tot-ASE to actor 2 and zero to actor 1 and the Planner, and r-SSE-ICC attributes r-SSE to four state variables matching the time-steps where actor 2 traverses colored cells.
- In the Sepsis example, 45.6% of TCFE is attributed to tot-ASE (about 0.374) and the remaining 54.4% to the state-specific part (about -0.446).
- Across 8728 selected alternative actions in Sepsis, the sum of individual agent-specific effects differs from the total agent-specific effect by up to 95%, while ASE-SV always attributes efficiently.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Administer vasopressors instead of mechanical ventilation at step 10 | Clinician in the sepsis scenario · Sepsis | successful treatment likelihood: 82% likelihood | Helped |
| Decompose the counterfactual effect into agent-mediated and state-mediated components | Causal explanation formula · Sepsis | TCFE attribution: 45.6% (tot-ASE ≈0.374); 54.4% (-r-SSE ≈0.446) | Mixed |
| Use the sum of individual clinician-specific and AI-specific effects instead of ASE-SV attribution | Agent-specific effect attribution · Sepsis | difference from total agent-specific effect: up to 95% | Hurt |
| Use ASE-SV to allocate total agent-specific effect between clinician and AI | Sepsis treatment agents · Sepsis | allocation efficiency | Helped |
| Increase clinician trust in the AI recommendations | ASE-SV agent attribution · Sepsis | clinician-attributed share of total agent-specific effect | Hurt |
| Increase clinician trust in the AI recommendations | ASE-SV agent attribution · Sepsis | AI-attributed share of total agent-specific effect | Helped |
Surprising findings
- TCFE is not in general decomposed into tot-ASE and SSE, a counter-intuitive result motivating the reverse SSE.
- The sum of clinician-specific and AI-specific effects can differ from the total agent-specific effect by up to 95%.
Limitations
- Relies on counterfactual identifiability assumptions (independent noise and weak noise monotonicity), which may be violated in practice.
- Exact ASE-SV scales exponentially with the number of agents, and r-SSE-ICC scales linearly with horizon.
- Each counterfactual sample requires LLM inference, which dominates computation time in the Gridworld experiment.
Experimental setup
Datasets: Gridworld environment (self-constructed) Sepsis management simulator
Benchmarks: Gridworld with LLM-assisted RL agents Sepsis management simulator
Metrics: total counterfactual effect total agent-specific effect reverse state-specific effect Shapley-based agent scores Gini coefficient of state contribution scores
Builds on: agent-specific effects intrinsic causal contributions MMDP-SCM causal mediation formula noise monotonicity
Closest papers by meaning
- Agent-Specific Effects: A Causal Effect Propagation Analysis in Multi-Agent MDPs (ICML 2024 · similarity 0.90)
- Collective Counterfactual Explanations: Balancing Individual Goals and Collective Dynamics (NeurIPS 2025 · similarity 0.86)
- Counterfactual Realizability (ICLR 2025 · similarity 0.86)
- Cyclic Counterfactuals under Shift–Scale Interventions (NeurIPS 2025 · similarity 0.85)
- Explaining Reinforcement Learning Agents through Counterfactual Action Outcomes (AAAI 2024 · similarity 0.85)
- Understanding Individual Agent Importance in Multi-Agent System via Counterfactual Reasoning (AAAI 2025 · similarity 0.85)
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.