This review systematically analyzes Reinforcement Learning approaches for self-healing in energy-constrained secure edge IoT networks across 82 studies from 2020 to 2026. Unlike existing surveys that focus on general RL applications, the proposed review focuses on a three-level taxonomy that uniquely addresses edge IoT deployment realities through formulation-scope-hardware mapping. The work develops a novel three-level taxonomy classifying recovery scope (node, link, service, network), RL formulations (tabular, deep, multi-agent, model-based), and constraint integration (energy, latency, security, hybrid), revealing service migration dominance at 30% coverage and node recovery achieving 38% maximum energy savings. Normalized performance baselines establish energy gains up to 44%, latency compliance of 84% under mobility traces, and 35% security exposure reduction during failover windows. 10 evidence-based gaps emerge, including a complete absence of model-based node recovery and multi-agent network security orchestration spanning only 2 papers. 15 prioritized future directions target 70% sample efficiency gains, 35% exposure reduction under compromised agents, and 22% Pareto improvements through joint constraint optimization, providing researchers and practitioners structured roadmap for sustainable edge IoT resilience. Performance metrics are normalized against static policy baselines using logarithmic scaling and success ratios to ensure cross-study comparability.
Hitesh Mohapatra (Thu,) studied this question.