Using Chinese dark chess endgame databases to validate and fine-tune game evaluation functions¹

Issue title: Chinese Dark Chess

Guest editors: Shun-Chin Hsu

Article type: Research Article

Authors: Chang, Hung-Jui^{a; b; *; **} | Chen, Jr-Chang^{c; ***} | Fan, Gang-Yu^{a; *} | Hsueh, Chih-Wen^{b; **} | Hsu, Tsan-sheng^{a; *; ****}

Affiliations: [a] Academia Sinica, Taiwan | [b] National Taiwan University, Taiwan | [c] National Taipei University, Taiwan

Correspondence: [****] Corresponding author. E-mail: tshsu@iis.sinica.edu.tw.

Note: [1] A preliminary earlier version of this paper appears in: H.-J. Chang, G.-Y. Fan, J.-C. Chen, C.-W. Hsueh and T.-s. Hsu. “Validating and Fine-tuning of Game Evaluating Functions using Endgame Databases.” In the 6th Computer Games Workshop at International Joint Conference on Artificial Intelligence 2017.

Note: [*] Institute of Information Science, Academia Sinica, Taipei 115, Taiwan.

Note: [**] Department of Computer Science and Information Engineering, National Taiwan University, Taipei 106, Taiwan.

Note: [***] Department of Computer Science and Information Engineering, National Taipei University, New Taipei City 237, Taiwan.

Abstract: Alpha–beta based game playing programs usually contain an evaluation function f to assess the value of every legal position by considering the location and material values of each piece on the board. Normally, a manually designed f is imperfect and requires a great amount of expert knowledge. Theoretically speaking, f can be a giant table which gives value for every position. However, this table has too many entries. Hence it takes huge space and is also time-consuming to construct. Note that when the number of pieces left on the board is small, say 6 for Chinese dark chess, such a giant table already exists and is called an endgame database. Endgame databases provide game-theoretical, i.e., perfect, values for all legal positions in the included endgame. Due to resource constraints, only a very selected number of endgame databases are built. Observing the strength of having perfect information on selected positions and the limitation of not being able to have information on all positions by using endgame databases to derive f, we propose a scheme to better design f. First, an f is manually designed, and then is validated and fine-tuned using abstracted critical information from existing endgame databases. In the process, we begin by validating and correcting f on positions when they are contained in endgame databases. Using this information, we then discover meta-knowledge to fine-tune and revise f so that f better evaluates a position, even if it is not in any given endgame database. The experiment results show that our approach is successful.

DOI: 10.3233/ICG-180048

Journal: ICGA Journal, vol. 40, no. 2, pp. 45-60, 2018