Javascript must be enabled for the correct page display

Reinforcement Learning in the Game Lines of Action: Input Representations and Look-ahead

Sasso, Remo and van Lohuizen, Quintin (2018) Reinforcement Learning in the Game Lines of Action: Input Representations and Look-ahead. Bachelor's Thesis, Artificial Intelligence.


Download (732kB) | Preview
[img] Text
Restricted to Registered users only

Download (94kB)


This paper investigates the application of reinforcement learning to the game Lines of Action by using a multi-layer perceptron (MLP). It investigates which training and testing method leads the agent to perform best, while applying a temporal difference learning algorithm. Three opponents are created in order to determine the most suitable training opponent: a random opponent, a fixed opponent and the agent itself. Additionally, different kinds of input representations are fed into the multi-layer perceptron, to see whether they affect the performance if used during training. Lastly, a testing method is introduced where the agent uses a look-ahead strategy. This allows the agent to perform a deep search, which may affect its performance. For this research the temporal difference learning method TD(0) was used in combination with an MLP. By using the training opponents as testing opponents, the resulting performances showed that the agent learns best against itself. The look-ahead play was tested in an identical manner, which produced a significant improvement in performance against opponents that are not completely random. Finally, we found that by training the agent with different game state representations, performance significantly increases when trained against a fixed opponent or the random opponent.

Item Type: Thesis (Bachelor's Thesis)
Supervisor name: Wiering, M.A.
Degree programme: Artificial Intelligence
Thesis type: Bachelor's Thesis
Language: English
Date Deposited: 28 Jul 2018
Last Modified: 30 Jul 2018 14:06

Actions (login required)

View Item View Item