Causal-intervention and uncertainty-aware resilience enhancement for multi-agent autonomous systems

Multi-agent autonomous systems must maintain mission performance in the presence of node failures and structural disruption. Because recovery operations alter functional dependencies, representations learned from nominal graphs can become inadequate for the states encountered during reconfiguration. We present causal-intervention and uncertainty-aware resilience enhancement (CURE), a two-phase framework that combines structural interventions specified by the simulator with uncertainty-aware multiobjective value decomposition. In Phase Ⅰ, a shared encoder that models relational structure is trained on nominal graphs and graphs containing node replacements or deletions, using predefined functional metapaths. The interventions modify node availability and functional assignments within the simulator. Regenerated path targets, together with shared supervision over relations, expose the encoder to their structural consequences. In Phase Ⅱ, the encoder is frozen, and recurrent local value functions shared across node types are learned with a monotonic value mixer under a budget of one operation per decision. An ensemble of Gaussian predictors provides estimates of aleatoric variance and disagreement among ensemble members. Analytical deficit weights distribute emphasis among sustained performance, rapid recovery, and steady-state performance, whereas predictive uncertainty parameterizes a log-sum-exp reward aggregation that increasingly prioritizes lower-performing objectives as confidence declines. On a heterogeneous swarm simulator, CURE achieves greater resilience than six reinforcement-learning baselines across eight attack scenarios at 50% damage. Additional experiments characterize performance for damage ratios from 30% to 80%, recovery speed, and variation across random topologies.