Learning Partial Action Replacement in Offline MARL

Dimensions overall score 7.0

GitHub Code Pulse

No public code linked for this paper yet.

• PLCQL reduces the number of per-iteration Q-function evaluations from n to 1 compared to the previous PAR-based method SPaCQL.technical 95%
  Evidencepartial

  Compared with the previous PAR-based method SPaCQL, PLCQL reduces the number of per-iteration Q-function evaluations from n to 1

  Implicationpartial

  Directly and explicitly stated in the abstract with a clear numeric comparison.

  Verificationpartial

  partial

• PLCQL achieves the highest normalised scores on 66% of tasks across MPE, MaMuJoCo, and SMAC benchmarks.result 90%
  Evidencepartial

  Empirically, PLCQL achieves the highest normalised scores on 66% of tasks across MPE, MaMuJoCo, and SMAC benchmarks

  Implicationpartial

  Explicitly stated in the abstract with a specific percentage and benchmark list.

  Verificationpartial

  partial

• PLCQL outperforms SPaCQL on 84% of tasks.result 90%
  Evidencepartial

  outperforming SPaCQL on 84% of tasks

  Implicationpartial

  Explicitly stated in the abstract with a specific percentage.

  Verificationpartial

  partial

• The estimation error of the proposed method scales linearly with the expected number of deviating agents.technical 85%
  Evidencepartial

  We prove a value-error bound showing that the estimation error scales linearly with the expected number of deviating agents.

  Implicationpartial

  Directly stated as a proven theorem in the abstract and analysis, though the full proof is not shown in the excerpt.

  Verificationpartial

  partial

• PLCQL is the first method to use a contextual bandit to adaptively select PAR subset sizes in offline MARL.method 80%
  Evidencepartial

  To our knowledge, PLCQL is the first method to use a contextual bandit to adaptively select PAR subset sizes in offline MARL.

  Implicationpartial

  Directly stated in the analysis excerpt, but the claim of being 'first' is a broader assertion that would require verification against all prior literature.

  Verificationpartial

  partial

• The PAR policy in PLCQL is learned using Proximal Policy Optimisation (PPO) with an uncertainty-weighted reward.method 95%
  Evidencepartial

  We introduce PLCQL, a framework that formulates PAR subset selection as a contextual bandit problem and learns a state-dependent PAR policy using Proximal Policy Optimisation with an uncertainty-weighted reward.

  Implicationpartial

  Directly and explicitly stated in both the abstract and the algorithm description section.

  Verificationpartial

  partial

• Existing PAR-based approaches (like SPaCQL) rely on computationally expensive enumeration of multiple subset configurations.limitation 85%
  Evidencepartial

  Partial Action Replacement (PAR) mitigates this by anchoring a subset of agents to dataset actions, but existing approach relies on enumerating multiple subset configurations at high computational cost

  Implicationpartial

  Directly stated as a motivation for the work in both the abstract and introduction.

  Verificationpartial

  partial

• The joint action space in offline MARL grows exponentially with the number of agents, making out-of-distribution (OOD) joint actions unavoidable.technical 90%
  Evidencepartial

  the joint action space grows exponentially with the number of agents, making dataset coverage exponentially sparse and out-of-distribution (OOD) joint actions unavoidable.

  Implicationpartial

  Presented as a fundamental, established challenge in the field and is the core motivation for the paper, directly stated in the abstract and introduction.

  Verificationpartial

  partial

Startup potential card