• Corpus ID: 14213233

Leandro Soriano Marcolino Advisor : Hitoshi Matsubara Multi-Agent Monte Carlo Go

@inproceedings{Marcolino2011LeandroSM,
  title={Leandro Soriano Marcolino Advisor : Hitoshi Matsubara Multi-Agent Monte Carlo Go},
  author={Leandro Soriano Marcolino and Hitoshi Matsubara and Carlo Go},
  year={2011}
}

• L. Marcolino, H. Matsubara, C. Go
• Published 2011
• Computer Science

Go is a strategic board game that is considered one of the greatest challenges for Arti cial Intelligence. Many algorithms have been proposed, trying to tackle this problem, but generally all of them generated players that could be easily defeated by a strong human opponent. UCT Monte Carlo Go is one of the most successful algorithms. The basic idea is to associate a tree search with pseudo-random simulations, used to evaluate the leaves. Nowadays, the literature is more focused on how to parallelize the UCT algorithm, producing a stronger player by increasing the computational power. However, there is a limit in the improvement that can be obtained. In this thesis we focus on how to improve the algorithm itself, producing a stronger player with the same amount of computation. We propose a Multi-Agent version of UCT Monte Carlo Go. The emergent behavior of a great number of agents have been successfully applied in the literature to deal with a great variety of problems. In this thesis, we use emergence and diversity to increase the quality of the Monte Carlo simulations, increasing the strength of the arti cial player as a whole. Instead of one agent playing against itself, di erent agents play in the simulation phase of the algorithm, leading to a better exploration of the search space. However, we found that using all possible agents leads to a weaker player. Therefore, we developed two learning algorithms, in order to nd a set of agents that can e ectively generate a strong player. The rst learning algorithm is a simple hill-climbing approach, it tries each agent one time and the agents that can improve the solution remain. The second learning algorithm is a simulated annealing approach. At each iteration, it decides randomly if it is going to add a new agent, or remove one from the database. The agent to be added/removed is selected randomly. Modi cations that decrease the current solution might also be accepted. With either learning algorithm, we could signi cantly overcome Fuego, a top Computer Go software. Emergence seems to be the next step of Computer Go development.