Using reinforcement learning in the context of computer games

Studenteropgave: Kandidatspeciale og HD afgangsprojekt

  • Ole Kaae Thomsen
4. semester, Datalogi, Kandidat (Kandidatuddannelse)
In this Thesis I investigate the ability of an agent, implemented using reinforcement learning, to learn and to adapt to a changing environment. I do this using two different methods of reinforcement learning, basic Q learning and a hierarchical method, MaxQ learning. In order to test the agent implemented using these methods, I first design Flag Hunter, a simple turnbased game that forms the basis for the testing of the agent in different situations. Flag Hunter requires the agent to go to the opponent''s base, pick up its flag, and return it to the agent''s homebase. In order to test the agents ability to learn, they were trained, first by playing the game without an opponent, and second, to play the game against opponents of different levels of randomness. It was found that the agent using the MaxQ method, would reach the goal significantly faster than the basic Q learning, but converge at about the same rate. The reasons for this was found and discussed. To examine the agents ability to adapt to a changing situation, the agents were trained without an opponent, and then set to play against the opponent. It was found that neither was very succesful at adapting.
Udgivelsesdatojun. 2004
ID: 61061601