Reinforcement Learning Enables Safe Engine Control

Researchers present a toolchain applying Deep Deterministic Policy Gradient (DDPG) reinforcement learning to transient load control on a single-cylinder HCCI engine testbench, demonstrated in an arXiv preprint (v1 Jan 28, 2025; v2 Feb 24, 2026). They incorporate real-time k-nearest-neighbor safety monitoring to avoid damaging pressure-rise events and achieve 0.1374 bar RMSE for indicated mean effective pressure. The agent also adapts to higher ethanol energy shares while maintaining safety.