Browser/Server based Experimental Environment for Reinforcement Learning

This paper has proposed an implement method using browser and server framework to train the reinforcement learning based AI (Artificial Intelligence). By taking advantage of the WebSocket to transfer information during the AI learning, the browser is in charge of rendering the experimental environment by HTML5 and JavaScript, and the server is realized by c++ to train the policy gradient AI mode which feeds the output actions back to its input layer. With the proposed framework, the rendering environment and the training environment can be implemented independently to different runtime environments. Finally, the effectiveness of the B/S based AI learning model proposed in this paper is validated through the application of the game `Flappy Bird'.