Tristan Cazenave

Computer Go: an AI oriented survey

Since the beginning of AI, mind games have been studied as relevant application fields. Nowadays, some programs are better than human players in most classical games. Their results highlight the efficiency of AI methods that are now quite standard. Such methods are very useful to Go programs, but they do not enable a strong Go program to be built. The problems related to Computer Go require new AI problem solving methods. Given the great number of problems and the diversity of possible solutions, Computer Go is an attractive research domain for AI. Prospective methods of programming the game of Go will probably be of interest in other domains as well. The goal of this paper is to present Computer Go by showing the links between existing studies on Computer Go and different AI related domains: evaluation function, heuristic search, machine learning, automatic knowledge generation, mathematical morphology and cognitive science. In addition, this paper describes both the practical aspects of Go programming, such as program optimization, and various theoretical aspects such as combinatorial game theory, mathematical morphology, and Monte Carlo methods. . 2001 Elsevier Science B.V. All rights.

Combining UCT and Nested Monte Carlo Search for Single-Player General Game Playing

Monte Carlo tree search (MCTS) has been recently very successful for game playing, particularly for games where the evaluation of a state is difficult to compute, such as Go or General Games. We compare nested Monte Carlo (NMC) search, upper confidence bounds for trees (UCT-T), UCT with transposition tables (UCT+T), and a simple combination of NMC and UCT+T (MAX) on single-player games of the past General Game Playing (GGP) competitions. We show that transposition tables improve UCT and that MAX is the best of these four algorithms. Using UCT+T, the program Ary won the 2009 GGP competition. MAX and NMC are slight improvements over this 2009 version.

A Parallel General Game Player

We have parallelized our general game player Ary on a cluster of computers. We propose multiple parallelization algorithms. For the sake of simplicity all our algorithms have processes that run independently and that join their results at the end of the thinking time in order to choose a move. Parallelization works very well for checkers, quite well for other two player sequential move games and not at all for a few other games.

A phantom-go program

This paper discusses the intricacies of a Phantom-Go program. It is based on a Monte-Carlo approach. The program called Illusion plays Phantom Go at an intermediate level. The emphasis is on strategies, tactical search, and specialized knowledge. The paper provides a better understanding of the fundamentals of Monte-Carlo search in Go.

Optimization of the nested Monte-Carlo algorithm on the traveling salesman problem with time windows

The traveling salesman problem with time windows is known to be a really difficult benchmark for optimization algorithms. In this paper, we are interested in the minimization of the travel cost. To solve this problem, we propose to use the nested Monte-Carlo algorithm combined with a Self-Adaptation Evolution Strategy. We compare the efficiency of several fitness functions. We show that with our technique we can reach the state of the art solutions for a lot of problems in a short period of time.

Abstract Proof Search

In complex games with a large branching factor such as Go, programs usually use highly selective search methods, heuristically expanding just a few plausible moves in each position. As in early Chess programs, these methods have shortcomings, they often neglect good moves or overlook a refutation. We propose a safe method to select the interesting moves using game definition functions. This method has multiple advantages over basic alphabeta search: it solves more problems, the answers it finds are always correct, it solves problems faster and with less nodes, and it is more simple to program than usual heuristic methods. The only small drawback is the requirement for an abstract analysis of the game. This could be avoided by keeping track of the intersections tested during the search, maybe with a loss of efficacy but with a gain in generality. We give examples and experimental results for the capture game, an important sub-game of the game of Go. The principles underlying the method are not specific to the capture game. The method can also be used with different search algorithms. This algorithm is important for every Go programmer, and is likely to interest other game programmers.

A Parallel Monte-Carlo Tree Search Algorithm

Monte-Carlo Tree Search is a powerful paradigm for the game of Go. In this contribution we present a parallel Master-Slave algorithm for Monte-Carlo Tree Search and test it on a network of computers using various configurations: from 12,500 to 100,000 playouts, from 1 to 64 slaves, and from 1 to 16 computers. On our own architecture we obtain a speedup of 14 for 16 slaves. With a single slave and five seconds per move our algorithm scores 40.5% against GNU Go , with sixteen slaves and five seconds per move it scores 70.5%. At the end we give the potential speedups of our algorithm for various playout times.

Score bounded Monte-Carlo tree search

Monte-Carlo Tree Search (MCTS) is a successful algorithm used in many state of the art game engines. We propose to improve a MCTS solver when a game has more than two outcomes. It is for example the case in games that can end in draw positions. In this case it improves significantly a MCTS solver to take into account bounds on the possible scores of a node in order to select the nodes to explore. We apply our algorithm to solving Seki in the game of Go and to Connect Four.

A Generalized Threats Search Algorithm

A new algorithm based on threat analysis is proposed. It can model existing related algorithms such as Lambda Search and Abstract Proof Search. It solves 6x6 AtariGo much faster than previous algorithms. It can be used in other games. Theoretical and experimental comparisons with other related algorithms are given.

Solving breakthrough with Race Patterns and Job-Level Proof Number Search

breakthrough is a recent race-based board game usually played on a 8×8 board. We describe a method to solve 6×5 boards based on (1) race patterns and (2) an extension of (JLPNS).