Learning to Perform a Tetris with Deep Reinforcement Learning

Video games and simulated environments have been a popular testing ground for many recent RL algorithms because of their speed, repeatability and scalability. However many of the state-of-the-art algorithms still fail at games requiring long-term planning.

In this work I train an agent to consistently perform the eponymous Tetris - clearing four lines at once - which requires planning a sequence of moves many pieces ahead rather than greedily clearing whatever is available now.