Application of the killer-tree heuristic and the lambda-search method to lines of action
Introduction
In the studies on shogi endgames, major progress has been made on the mating-search technique using λ1-search [17]. In addition, related to the mating search, a smart technique for reusing the results of previous mating searches was proposed by Tanase [16], which is generalized as the killer-tree heuristic (KTH) [14]. These two search techniques have not yet well been studied in other domains than shogi. This paper explores the possibility of applying these techniques in the domain of lines of action (LOA).
LOA is a two-person connection-based game with perfect information. Since the aim of a player is to connect all his stones, important characteristics of the game are: concentration, centralization, solid formations, connections and (partial) blocking. They are very different from those in chess or shogi. A further noteworthy difference is the complexity: the state-space complexity and game-tree complexity of LOA are much lower than those for shogi or chess [18]. Therefore, in order to develop a strong LOA program, we probably need other methods than for chess or shogi. In the domain of computer shogi, several endgame-search techniques (e.g., PN*-search [15]) and its enhancements have been developed with significant results. In this paper we investigate the research question whether the effectiveness of these search techniques also holds for the endgame search for a connection-based game like LOA.
In Section 2 we will discuss the above-mentioned endgame techniques: lambda-search and the KTH. Then, in Section 3, we will briefly describe the game of LOA and give reasons why this domain is chosen as subject for our research. Section 4 then will describe the set-up of the experiments and will give experimental results and a discussion thereof. Finally, Section 5 will present our main conclusions.
Section snippets
Endgame-search techniques
In this section we describe the two endgame-search techniques investigated in this paper: lambda-search and the KTH.
Endgame search in lines of action
LOA is a connection-based two-player game with perfect information using an 8 × 8 board. It has been invented by Claude Soucie around 1960 [11]. The basic features for developing a program for LOA are detailed in [18], [20].
Experimental results and discussions
In this section we describe the set-up of the experiments and present and discuss all results.
Conclusions
In the present contribution, we have examined two search techniques in the domain of LOA.
The first technique is λ1-search. The use of pure λ1-search appears to be quite ineffective to find a winning sequence, because many solutions contain non-λ1-moves. The λ1-sorting method, in which λ1-moves are sorted in the front of a search, reduces the number of nodes searched roughly by half. However, the solving ability is declined because of the inefficiency of the λ1-move generation. This could be
Acknowledgements
We thank Mark Winands for kindly offering us his LOA test set. We further greatly acknowledge the constructive comments of three anonymous referees on an earlier version of this paper.
References (20)
- et al.
Computer shogi
Artificial Intelligence
(2002) - et al.
The PN*-search algorithm: application to tsume-shogi
Artificial Intelligence
(2001) - et al.
The principal continuation and the killer heuristic
- et al.
A brinkmate solver and its evaluation
Transactions of the Japanese Society for Artificial Intelligence
(2001) Using similar positions to search game trees
- et al.
Natural developments in game research
ICCA Journal
(1996) A new AND/OR tree search algorithm using proof number and disproof number
- A. Nagai, A new depth-first-search algorithm for AND/OR trees, M.Sc. Thesis, Department of Information Science, The...
- A. Nagai, Df-pn Algorithm for Searching AND/OR Trees and Its Applications, Ph.D. Thesis, Department of Information...
- A. Nagai, H. Imai, Application of df-pn+ to Othello endgames, in: Proceedings of Game Programming Workshop in Japan...
Cited by (7)
Game 5.0: social cognitionparallel game based on the parallel systems and machine game
2021, Chinese Journal of Intelligent Science and TechnologyMonte Carlo Game Solver
2020, arXivHierarchical Task Network Plan Reuse for video games
2016, IEEE Conference on Computatonal Intelligence and Games, CIGGame-tree search using proof numbers: The first twenty years
2012, ICGA JournalEvaluation-function based proof-number search
2011, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)Monte Carlo tree search in lines of action
2010, IEEE Transactions on Computational Intelligence and AI in Games