Obstacles Avoidance with Machine Learning Control Methods in Flappy Birds Setting

Object avoidance is an important topic in control theory. Various traditional control methods can be applied to achieve control of object path such as PID, Bang-Bang control, sliding mode control. Given a known and simple dynamic system, those classical controls method work pretty well. However, controls of complex, multi degree of freedom systems or controls of systems with unknown dynamics have been pushing the limit of traditional control laws. This report adopts machining learning methods of Support Vector Machine(SVM) with linear kernels and reinforcement learning using value iteration to solve control problems in the game ‘Flappy Bird’ without understanding the dynamics of the problem. For comparison purposes, Bang-Bang control is also implemented and described in the report. The game is also modified to increase difficulty for further comparison. The machine learning methods are shown to significantly improve the results got from BangBang Control. In the modified game version with moving pipes, Reinforcement Learning is more feasible than SVM. Detailed implementation of and comparison among each methods are discussed in this report.