Reijer Grimbergen

Natural Developments in Game Research

In game-programmmg research there are four Interestmg and related domams CHESS, XIANG QI (Chmese chess), SHOGI (Japanese chess) and Go In this article we compare CHESS With SHOGI by rules and by computatIOnal aspects We will see that CHESS and SHOGI are mostly very Similar, but that there are some Important differences which complicate SHOG! programmmg The most Important difference IS the game-tree compleXity, which IS considerably higher than the game-tree compleXity of CHESS We will then argue that the SimilarIties and differences make SHOGI a good chOice for advanced research m game programmmg In the near future CHESS Will no longer be competitively mterestmg Smce XIANG QI has a game-tree compleXity Similar to CHESS, the same AI techmques will also be successful In this domam, and, as a consequence, this game Will also no longer be mterestmg Go IS too rIsky as a next research target because lIttle IS known about the cogmtIve aspects of the game, which m our view hold the key to developmg new techmques A short history of computer Shogl With the results of the latest CSA computer Shogl tournament IS given ConclUSIOns are proVided m Section 5 In the appendiX, a short mtroductlOn to the rules of the game IS mcluded

A Shogi Program Based on Monte-Carlo Tree Search

Recently, Monte-Carlo Tree Search (MCTS) has been attracting a great deal of attention in gameprogramming research. This method has been quite successful in computer Go, but so far results in other games have not been so impressive. In this paper, we present an implementation of MCTS in shogi which combines techniques used in computer Go with a number of shogi-specific enhancements. We tested this implementation on a standard test set of tactical positions. The number of correct answers indicates that the strength of the Monte-Carlo based shogi program is about that of a 1-dan amateur. The results did not carry over to actual playing strength as the match results against a conventional shogi program with a strength of about a 1-dan player showed. Therefore, it seems unlikely that a pure MCTS-based shogi program will surpass the level of the best conventional shogi programs. However, we also observed that our MCTS program could solve certain opening and endgame positions that are considered hard to solve with the current methods. Therefore, we believe that MCTS can be a useful method to improve the overall performance of a shogi program.

A Survey of Tsume-Shogi Programs Using Variable-Depth Search

Recently, a number of programs have been developed that successfully apply variable-depth search to find solutions for mating problems in Japanese chess, called tsume shogi. Publications on this research domain have been written mainly in Japanese. To present the findings of this research to a wider audience, we compare six different tsume programs. To find the solutions of difficult tsume-shogi problems with solution sequences longer than 20 plies, we will see that variable-depth search and hashing to deal with a combination of transposition, domination and simulation leads to strong tsume-shogi programs that outperform human experts, both in speed and in the number of problems for which the solution can be found. The best program has been able to solve Microcosmos, a tsume-shogi problem with a solution sequence of 1525 plies.

Plausible Move Generation Using Move Merit Analysis with Cut-Off Thresholds in Shogi

In games where the number of legal moves is too high, it is not possible to do full-width search to a depth sufficient for good play. Plausible move generation (PMG) is an important search alternative in such domains. In this paper we propose a new method for plausible move generation in shogi. During move generation, Move Merit Analysis (MMA) gives a value to each move based on the plausible move generator(s) that generated the move. These values can be used for different cut-off schemes. We investigate the following alternatives: 1) Keep all moves with a positive MMA value; 2) Order the moves according to their MMA value and use cut-off thresholds to keep the best N moves. PMG with MMA and cut-off thresholds can save between 46% and 68% of the total number of legal moves with an accuracy between 99% and 93%. Tests show that all versions of shogi programs using PMG with MMA outperform an equivalent shogi program using full-width search. It is also shown that MMA is vital for our approach. Plausible move generation with MMA performs much better than plausible move generation without MMA. Cut-off thresholds improve the performance for N = 20 or N = 30.

Enhancing search efficiency by using move categorization based on game progress in amazons

Amazons is a two-player perfect information game with a high branching factor, particularly in the opening. Therefore, improving the efficiency of the search is important for improving the playing strength of an Amazons program. In this paper we propose a new method for improving search in Amazons by using move categories to order moves. The move order is decided by the likelihood of the move actually being selected as the best move. Furthermore, it will be shown that the likelihood of move selection strongly depends upon the stage of the game. Therefore, our method is further refined by adjusting the likelihood of moves according to the progress of the game. Self-play experiments show that using move categories significantly improves the strength of an Amazons program and that combining move categories with game progress is better than using only move categories.

Review: Computer Shogi through 2000

Since the first computer shogi program was developed by the first author in 1974, more than a quarter century has passed. During that time, shogi programming has attracted both researchers and commercial programmers and playing strength has improved steadily. Currently, the best programs have a level that is comparable to that of a strong amateur player (about 4-dan), but the level of experts is still beyond the horizon. The basic structure of strong shogi programs is similar to chess programs. However, the differences between chess and shogi have led to the development of some shogi-specific methods. In this paper we will give an overview of the computer shogi history, summarise the most successful techniques and give some ideas for the future directions of research in computer shogi.

A Shogi Processor with a Field Programmable Gate Array

In this paper we describe the architecture of a shogi processor based on reconfigurable hardware. For our implementation, we have used Field Programmable Gate Arrays (FPGAs), which can be reconfigured dynamically by downloading configuration data from host computers. Because of this reconfiguration flexibility, it is possible to implement and evaluate new algorithms quickly and to make small subsystems (of very low cost) that can be used on demand. For shogi these two features are especially important, as there are no stable subsystems that can be ported to special purpose hardware. Also, in shogi different modules are needed for different stages of the game. To test the feasibility of using FPGAs for shogi, we have implemented two modules that are general for all strong shogi programs on one off-the-shelf PCI board with one FPGA. The piece cover module on an FPGA is 62 times faster than the software module, while the module for finding mate on an FPGA is 9 times faster than the software module.

Cognitive Modeling of Knowledge-Guided Information Acquisition in Games

Since Chase and Simon presented their influential paper on perception in chess in 1973, the use of chunks has become the subject of a number of studies into the cognitive behavior of human game players. However, the nature of chunks has remained elusive, and the reason for this lies in the lack of using a general cognitive theory to explain the nature of chunks. In this paper it will be argued that Marvin Minskys Society of Mindtheory is a good candidate for a cognitive theory to define chunks and to explain the relation between chunks and problem-solving tasks. To use Minskys Society of Mindtheory to model human cognitive behavior in games, we first need to understand more about the primitive agents dealing with the relation between perception and knowledge in memory. To investigate this relation, a reproduction experiment is performed in shogi showing that perception is guided by knowledge in long-term memory. From the results we may conclude that the primitive agents in a cognitive model for game-playing should represent abstract concepts such as board, piece, and kingrather than the perceptual features of board and pieces.

Board Maps and Hill-Climbing for Opening and Middle Game Play in Shogi

Most strong game-playing programs use large, well tuned opening books to guide them through the early stages of the game. However, in shogi (Japanese chess) the classic approach of building a large opening book of known positions is infeasible. In this paper, we present a different approach for opening and middle game play in shogi. This method uses board maps that assign values to each square for each piece in a number of different formations. Hill-climbing is then used to guide pieces to optimal squares. We define board maps for defensive piece formations (castles), attacking formations (assaults) and for recognizing the type of opening position. Results show that using board maps in combination with hill-climbing significantly improves the playing strength of a shogi program. Furthermore, using maps for both castles and assaults is better than using only maps for castles.

Pattern Analysis and Analogy in Shogi: Predicting Shogi Moves from Prior Experience

Abstract. As a research paradigm, pattern analysis has been shown to be an effective tool for analyzing complex game situations in both chess and go. We extend the prior pattern analysis research in chess to the domain of shogi. Shogi is computationally more complex than chess and should realize greater benefits than the chess domain from pattern recognition and pattern exploitation research. The IAM program, which has accurately predicted up to 28% of the moves for a specific chess player, is redesigned to operate in the domain of shogi. Results similar to those achieved for the domain of chess are achieved in shogi.