Research map / Multi-agent systems and games / Games and equilibria
Agent-Specific Effects: A Causal Effect Propagation Analysis in Multi-Agent MDPs
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.
Introduces agent-specific effects (ASE) and counterfactual ASE (cf-ASE) for multi-agent MDPs, gives identifiability under noise monotonicity, a sampling-based estimator, and evaluates on Graph and Sepsis simulators.
Problem
Quantifying how much of an agent's action's causal/counterfactual effect on an outcome propagates through other agents' responses in multi-agent sequential decision-making (e.g., AI-clinician sepsis treatment), where existing path-specific effects cannot express this.
Approach
Defines agent-specific effects as a three-world natural-intervention quantity in which an selected set of effect agents respond to the intervened action while non-effect agents keep factual actions; shows cf-ASE is identifiable under a weak noise monotonicity assumption (not requiring bijective models), and gives an abduction-action-prediction sampling estimator.
Key results
- On the Graph environment, cf-ASE estimates stay accurate for small policy estimation error (epsilon_max <= 0.05), while wrong total orderings can significantly increase error.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase clinician trust in the AI (μ) | Sepsis clinician-specific effect measured by cf-ASE · Sepsis | clinician-specific effect | Hurt |
| Increase clinician trust in the AI (μ) | Sepsis AI-specific effect measured by cf-ASE · Sepsis | AI-specific effect | Helped |
| Increase clinician trust in the AI (μ) | Sepsis clinician-specific effect measured by cf-PSE · Sepsis | clinician-specific effect | No clear effect |
| Use cf-PSE instead of cf-ASE to measure clinician-specific effects | Sepsis · Sepsis | clinician-specific effect | Mixed |
| Use cf-PSE instead of cf-ASE to measure AI-specific effects | Sepsis · Sepsis, μ = 0 | AI-specific effect | Hurt |
| Estimate agent policies with small errors (εmax ≤ 0.05) | cf-ASE estimation · Graph | average absolute estimation error | No clear effect |
| Use an incorrect total ordering instead of the correct ordering | cf-ASE estimation · Graph | average absolute estimation error | Hurt |
| Assume the reverse of the correct total ordering | cf-ASE estimation · Graph | average absolute estimation error | No clear effect |
| Increase clinician trust in the AI (μ) | Sepsis clinician-specific cf-ASE action distribution · Sepsis | fraction of actions with clinician-specific effect > 0.75 | Hurt |
| Increase clinician trust in the AI (μ) | Sepsis clinician-specific cf-ASE action distribution · Sepsis | fraction of actions with clinician-specific effect < 0.25 | Helped |
| Increase clinician trust in the AI (μ) | Sepsis AI-specific cf-ASE action distribution · Sepsis | fraction of actions with AI-specific effect > 0.75 | Helped |
| Increase clinician trust in the AI (μ) | Sepsis AI-specific cf-ASE action distribution · Sepsis | fraction of actions with AI-specific effect < 0.25 | Hurt |
Surprising findings
- cf-PSE can assign positive clinician-specific effect even when the clinician blindly follows AI recommendations, a counter-intuitive result that cf-ASE avoids.
Limitations
- Identifiability requires noise monotonicity, which may be violated in practice; Sepsis experiments assume it without verification.
- Non-identifiability of ASE under the general setting and no treatment of unobserved confounding.
- Experiments use simulated environments only (Graph and Sepsis), and no causal explanation formula is developed yet.
Experimental setup
Datasets: Graph environment (6 agents) Sepsis simulator
Benchmarks: Graph Sepsis
Metrics: average absolute error of cf-ASE average cf-ASE value distribution of cf-ASE values
Baselines: total counterfactual effect (TCFE) counterfactual path-specific effect (cf-PSE)
Builds on: path-specific effects structural causal models counterfactual stability Policy Iteration
Closest papers by meaning
- Counterfactual Effect Decomposition in Multi-Agent Sequential Decision Making (ICML 2025 · similarity 0.90)
- Situation-Dependent Causal Influence-Based Cooperative Multi-Agent Reinforcement Learning (AAAI 2024 · similarity 0.84)
- Sim-to-Real Causal Transfer: A Metric Learning Approach to Causally-Aware Interaction Representations (CVPR 2025 · similarity 0.84)
- Higher-Order Causal Message Passing for Experimentation with Complex Interference (NeurIPS 2024 · similarity 0.84)
- A Principle of Targeted Intervention for Multi-Agent Reinforcement Learning (NeurIPS 2025 · similarity 0.84)
- Detecting Influence Structures in Multi-Agent Reinforcement Learning (ICML 2024 · similarity 0.83)
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.