Multi-Objective Constrained Reinforcement Learning for Joint Routing–MAC–Duty Cycling in Low-Power Wireless Sensor Networks

[Purpose] The aim of this study is to introduce a Constrained Multi-Objective Reinforcement Learning Model (CMORLM) for optimizing Joint Routing, Medium Access Control (MAC), and Duty Cycling Optimization (DCO) in low-power WSNs. [Design/methodology/approach] In this paper, we have suggested the CMORLM approach as a constrained Markov Decision Process (MDP) with three competing goals: lowering Energy Consumption (EC), lowering End-to-End-Latency (EEL), and raising Packet Delivery Ratio (PDR). There are strict limits on the amount of residual energy, the buffer size, and the Quality of Service (QoS) requirements. Lagrangian Constraint Handling (LCH) and multi-objective policy gradients are combined within the primal-dual optimization method. For routing, MAC, and DCO, the policy network uses a shared encoder with factorized heads. Federated Gradient Aggregation (FGA) is used for distributed learning across Sensor Nodes (SN). [Findings] Testing in NS-3 shows that EC is 34.2% lower, EEL is 41.3% higher, and PDR is 16.5% higher than Traditional Layered Protocols (TLP). Network Lifetime (NL) goes up by 38.4%. The constraint violation rate (CVR) is still below 1%. Ablation studies show that joint optimization increases the EC by 44.7% over single-layer control. The suggested CMORLM works well on networks with 50 to 200 nodes and can handle changes in traffic, node failures, and mobile sinks. [Research implications/Originality/value] Pareto frontier analysis is performed to enable operator control over performance trade-offs through weight configuration.