Analysis on Deep Reinforcement Learning with Flappy Brid Gameplay

In recent years, machine learning has become a popular area in the IT field. In 2016, a game AI AlphaGo turned out to defeat many Go masters. The technology behind AlphaGo: the popularity of deep reinforcement learning is gradually increasing. This paper aims to use deep reinforcement learning algorithm to train a specific model which can play the Flappy Bird game autonomously. In this paper, two types of optimizers: Adam and RMSProp are used in the deep neural network during the training sessions and the testing results with the models are compared to figure out which model has a better performance when playing Flappy Bird. From the testing results, when the training rounds is insufficient, the model with Adam optimizer performs better than the model with RMSProp optimizer. However, when the training rounds are large enough, the performance of the model with RMSProp optimizer is almost 2 times better than the model with Adam optimizer. After comparison between the two models, this paper finds out that with the increasing of the training rounds, the performance of model with RMSProp optimizer will gradually exceed the model with Adam optimizer.