MAGIC estimates multi-step action effects between agents with counterfactual interventions, gates them by advantage, and converts them to intrinsic rewards, yielding 26.9% and 10.1% relative gains on MPE and SMAC benchmarks.
Advances in Neural Information Processing Systems , volume=
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.MA 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
MAGIC: Multi-Step Advantage-Gated Causal Influence for Multi-agent Reinforcement Learning
MAGIC estimates multi-step action effects between agents with counterfactual interventions, gates them by advantage, and converts them to intrinsic rewards, yielding 26.9% and 10.1% relative gains on MPE and SMAC benchmarks.