Investigation into the effect of social learning in reinforcement learning board game playing agents
Date
2009-09-14T10:13:47Z
Authors
Marivate, Vukosi Ntsakisi
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
This thesis presents the use of social learning to improve the performance of game
playing reinforcement learning agents. Agents are placed in a social learning environment
as opposed to the Self-Play learning environment. Their performance is monitored and
analysed in order to observe how the performance changes compared to Self-Play agents.
Two case studies were conducted, one with the game Tic-Tac-Toe and the other with the
African board game of Morabaraba. The Tic-Tac-Toe agents used a table based TD ( )
algorithm to learn the Q values. The results from the tests for the Tic-Tac-Toe agents
indicate that the social learning agents perform better than the Self-Play agents in both
board tests and competitive tests. By increasing the population sizes of the agents the
number of superior social agents also increases as well as improvements in their skill
level. In the second case study the agents use function approximation and the TD ( )
algorithm because of a larger number of states. The social agents performed better than
the Self-Play agents in the board tests and are not superior in the test where they compete
against each other. Larger populations were not possible with the Morabaraba agents but
the results are still positive as the agents perform well in the board tests.