With the increasing complexity of industrial process, accurate models of automatic control systems are difficult to be established, which makes model-based control methods become less effective. On this premise, a reinforcement learning (RL) based data-driven control method for dynamic systems is proposed in this paper. For our purposes, a model-free RL controller with an Actor-Critic structure is developed by using neural networks (NN), combining the merits of both value-based and policy-based RL algorithms. Policy of the actor is assessed by the critic, updating parameters of RL controller in real-time for online optimization. The effectiveness of the RL controller is demonstrated by an experimental study on FLUIDMechatronix platform.