Taxonomy and Benchmark of AI/ML-Based Energy-Efficient Routing Protocols for Wireless Sensor Networks

AI/ML-based routing protocols for Wireless Sensor Networks (WSNs) promise substantial energy gains, yet the field lacks a unified analytical framework for cross-paradigm comparison. Existing papers adopt incompatible simulation setups, report only communication-layer energy metrics, and rarely validate results on physical hardware. This paper develops a twelve-dimension taxonomy covering 40 published works across supervised learning, reinforcement learning, neuro-fuzzy, and evolutionary computation paradigms, and pairs it with a reproducible NS-3 v3.40 benchmark using standardised parameters. Four systematic research gaps emerge: no paper models joint communication-plus-computation energy; fewer than five report any hardware measurement; none evaluate generalisation across traffic domains; and all 40 withhold source code. Benchmark projections show that Q-learning achieves 30% first-node-death (FND) improvement over LEACH and ANFIS achieves 24% with 0.6% computation overhead, whereas DQN consumes approximately 95% of node energy per round on 8-bit hardware, negating its simulated gains. These findings provide the evidential basis for a companion paper series resolving each identified gap