Reinforcement Learning for Engine Idle Speed Control

A control method of neural network controller with reinforcement learning is proposed to implement idle speed control of an automobile engine to reduce fluctuation of the idle speed. Firstly, the reinforcement-learning neural network is demonstrated in detail. Then, the control scheme of the reinforcement-learning controller is designed to experiment. Q learning algorithm, as one of methods of reinforcement learning, is used for learning of the neural network, which is based on evaluating the system performance and giving credit for successful actions. After the proposed controller is trained fully, the contrast experiments are implemented on an engine test bench between the proposed controller and the original controller. Experimental results show that the reinforcement learning controller has better performance in speed fluctuation and its frequency and fuel economy than that of the original controller. And, the transition of the transient idle speed controlled by the proposed controller is more smooth and stable. Meanwhile, exhaust emissions are tested during the conditions controlled by the two types of controllers respectively. And results demonstrate that the proposed controller has better fuel economy because of its lower exhaust emissions.