The rise of deep learning has revolutionised the fields of machine learning and artificial intelligence in recent years. It enabled reinforcement learning – a machine learning approach focused on learning by trial-end-error – to scale to previously intractable high-dimensional problems, creating the field of deep reinforcement learning.
In deep reinforcement learning, it is common practice to use simulated environments in order to speed up training, and computer games are ideal candidates for this role. They not only provide challenging tasks for artificial agents, but also attract the attention of the public. Furthermore, they make the results directly comparable to that of a human’s, which is an important baseline in artificial intelligence research.
The goal of this thesis is the creation of an intelligent agent that can win the game of 2048, which is a once very popular stochastic puzzle game. The simple mechanics, yet challenging goals, make 2048 a suitable environment for stepping into the world of deep reinforcement learning. Using policy gradient algorithms, multiple agents were trained during the course of the development.
This work provides an in-depth review of the relevant literature, and thoroughly discusses the development of both the environment and the agents. The thesis also provides a detailed evaluation of the capabilities of the agents. Lastly, it gives the conclusion regarding the whole development process and the results of the training.