Improving a Proportional Integral Controller with Reinforcement Learning on a Throttle Valve Benchmark

• DOI: 10.48550/arxiv.2402.13654
• Publication Date: 2024-02-21
• Authors: Paul Daoudi; Bojan Mavkov; Bogdan Robu; Christophe Prieur; Emmanuel Witrant; Ludovic Dos Santos; Emmanuel Witrant; Merwan Barlier
• Keywords: Throttle; Benchmark (surveying); Sample (material)

This paper presents a learning-based control strategy for non-linear throttle valves with an asymmetric hysteresis, leading to near-optimal controller without requiring any prior knowledge about the environment. We start carefully tuned Proportional Integrator (PI) and exploit recent advances in Reinforcement Learning (RL) Guides improve closed-loop behavior by learning from additional interactions valve. test proposed method various scenarios on three different valves, all highlighting benefits of combining both PI RL frameworks performance stochastic systems. In experimental cases, resulting agent has better sample efficiency than traditional agents outperforms controller.