Supervisors info:
Επιβλέπων Καθηγητής Διονύσιος Ρεΐσης, Αναπληρωτής Καθηγητής, Τμήμα Φυσικής, ΕΚΠΑ
Δρ Νικόλαος Βλασσόπουλος, Επιστημονικός Συνεργάτης
Summary:
Summary The aim of this Master Thesis is to develop a Tower Defense game and to equip it with computational intelligence. A Deep Reinforcement Learning technique, more specifically the Advantage Actor-Critic (A2C), is implemented and tuned in order to provide the enemy with intelligence assisting him to reach the final goal by avoiding the obstacles.
The enemy is trained online. The A2C method, due to its actor-critic part, benefits from the characteristics of both policy search and Q value methods, while the use of a deep neural network permits the handling of large action spaces by reducing the dimension of the problem. The reinforcement learning is a trial-and-error general method letting the agent to learn via the rewards he is receiving from the environment in response to his actions.
By enriching the game with an artificial intelligence component, it is becoming more interesting for the player. In addition, the game served as a means to study and deepen our knowledge further in the field of deep reinforcement learning.
Chapter 2 is discussing aspects and principles of Machine Learning.
Chapter 3 covers the topic of Neural Networks by presenting their basic components, characteristics, learning/training methods and the most important architectures.
The subject of reinforcement learning is discussed in chapter 4. The fundamental terms, definitions and principles are addressed in the beginning of the chapter, then the basic reinforcement learning algorithms are briefly presented. Deep reinforcement learning and two relevant algorithms are covered at the end of the chapter.
Chapter 5 deals with the game, more specifically with the design, development, implementation and the reinforcement learning component. The approach followed and the design characteristics, the tools that were employed for the development and the issue of tuning the training parameters are presented in detail. The chapter ends with a presentation and discussion of the results obtained after running the game.
The thesis is finishing with the Conclusions, whereas the full code that has been developed is presented in the Appendix.
Keywords:
Reinforcement Learning, Neural Networks, A2C, Python, Pygame, TensorFlow, Keras, Tower Defense
