Evidence-Bounded Edge Reinforcement Learning for Adaptive Fuel Optimization

Abstract

Edge intelligence is frequently proposed for vehicle energy management, yet published performance claims are difficult to assess when training artifacts, scenario definitions, and hardware logs are unavailable. This study develops an evidence-bounded and fully reproducible evaluation of a continuous-action Proximal Policy Optimization (PPO) controller for adaptive fuel management, together with a heterogeneity-aware federated extension. The state comprises vehicle speed, acceleration, engine displacement, cylinder count, instantaneous fuel consumption, and carbon-dioxide rate; the action is bounded throttle; and the reward jointly penalizes fuel use, proportional CO₂, jerk, and speed-tracking error. After 50,000 training steps, a preregistered-style paired evaluation over 30 deterministic driving-cycle seeds reduced mean simulated fuel consumption from 4.1221 to 3.9134 L/100 km. The mean reduction was 5.06% (bootstrap 95% CI 5.04–5.08), and tracking error decreased by 37.15% (95% CI 37.06–37.23). Edge policy computation averaged 0.0042 ms on the development host, whereas an explicitly modeled 100 ms round trip yielded 100.0042 ms end-to-end decision latency. A ten-client, ten-round weighted federated experiment retained a 4.85% efficiency gain. The results establish software reproducibility and architectural latency implications, but do not claim real-vehicle validity: the source manuscript’s SUMO–CARLA traces, trained checkpoints, and hardware-in-loop logs were not present. This distinction yields a defensible benchmark and a concrete evidence protocol for subsequent automotive validation.