Online Resilience Recovery Scheduling for Unknown Topology Wireless Ad Hoc Networks Using Correlation Guided Combinatorial Bandits
Managing resilience in wireless ad hoc networks is challenging when the topology, routing paths, and protocol details are not fully observable. We study an online resilience recovery scheduling problem where a controller selects a limited set of nodes each round for monitoring, reconfiguration, protection, or recovery, and only an aggregate network-level feedback is observed. To address the resulting credit assignment issue in combinatorial decisions, we develop (i) an Improved Combinatorial UCB method that estimates node utilities from aggregate rewards, and (ii) a Correlation-Guided UCB method that learns a pairwise node correlation matrix and constructs more effective multi-node recovery sets. We also evaluate a protocolindependent reward based on active-node variation to support partially observable environments. Simulations on Poisson point process wireless ad hoc networks show that the proposed correlation-guided approach consistently improves cumulative service recovery and robustness across network density, size, flow load, and recovery budgets, outperforming independent bandits and centrality-based baselines.