Reproduce “On Blame Attribution for Accountable Multi-Agent Sequential Decision Making” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Stelios Triantafyllou, Adish Singla, Goran Radanovic
The paper examines blame allocation for inefficiency in sequential decision-making. It compares improvements in a multi-agent Markov decision process, including average participation, which divides coalition improvements among pivotal members. Average participation is valid but average-efficient, and its performance guarantee requires unchanged pivotality. In Gridworld simulations, it gives A2 minimum blame when its model matches A1’s policy, unlike Shapley and Banzhaf.
Key insights
- The desired accountability properties are fundamentally incompatible: no method can simultaneously guarantee efficiency, symmetry, invariance, and full performance monotonicity.
- Shapley value gives a complete and symmetric allocation but can violate coalition rationality and assign more blame after an agent improves performance; marginal contribution preserves performance monotonicity but can over-blame or miss coordination failures.
- Average participation trades some explanatory completeness for validity, symmetry, invariance, and stronger protection against over-blaming, while Blackstone consistency ensures that uncertainty does not increase an agent’s blame.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.