Wen-Jie Tseng

Job-Level Alpha-Beta Search

An approach called generic job-level (JL) search was proposed to solve computer game applications by dispatching jobs to remote workers for parallel processing. This paper applies JL search to alpha-beta search, and proposes a JL alpha-beta search (JL-ABS) algorithm based on a best-first search version of MTD(f). The JL-ABS algorithm is demonstrated by using it in an opening book analysis for Chinese chess. The experimental results demonstrated that JL-ABS reached a speed-up of 10.69 when using 16 workers in the JL system.

Job-level algorithms for Connect6 opening position analysis

This paper investigates job-level (JL) algorithms to analyze opening positions for Connect6. The opening position analysis is intended for opening book construction, which is not covered by this paper. In the past, JL proof-number search (JL-PNS) was successfully used to solve Connect6 positions. Using JL-PNS, many opening plays that lead to losses can be eliminated from consideration during the opening game. However, it is unclear how the information of unsolved positions can be exploited for opening book construction. For this issue, this paper first proposes four heuristic metrics when using JL-PNS to estimate move quality. This paper then proposes a JL upper confidence tree (JL-UCT) algorithm and some heuristic metrics, one of which is the number of nodes in each candidate moves subtree. In order to compare these metrics objectively, we proposed two kinds of measurement methods to analyze the suitability of these metrics when choosing best moves for a set of benchmark positions. The results show that for both metrics this node count heuristic metric for JL-UCT outperforms all the others, including the four for JL-PNS.

Strength Improvement and Analysis for an MCTS-Based Chinese Dark Chess Program

Monte-Carlo tree search (MCTS) has been successfully applied to Chinese dark chess (CDC). In this paper, we study how to improve and analyze the playing strength of an MCTS-based CDC program, named DarkKnight, which won the CDC tournament in the 17th Computer Olympiad. We incorporate the three recent techniques, early playout terminations, implicit minimax backups, and quality-based rewards, into the program. For early playout terminations, playouts end when reaching states with likely outcomes. Implicit minimax backups use heuristic evaluations to help guide selections of MCTS. Quality-based rewards adjust rewards based on online collected information. Our experiments showed that the win rates against the original DarkKnight were 60.75 %, 70.90 % and 59.00 %, respectively for incorporating the three techniques. By incorporating all together, we obtained a win rate of 76.70 %.

An analysis for strength improvement of an MCTS-based program playing Chinese dark chess

Monte Carlo tree search (MCTS) has been successfully applied to many games recently. Since then, many techniques are used to improve the strength of MCTS-based programs. This paper investigates four recent techniques: early playout terminations, implicit minimax backups, quality-based rewards and progressive bias. The strength improvements are analyzed by incorporating the techniques into an MCTS-based program, named DarkKnight, for Chinese Dark Chess. Experimental results showed that the win rates against the original DarkKnight were 60.75%, 71.85%, 59.00%, and 82.10%, respectively for incorporating the four techniques. The results indicated that the improvement by progressive bias was most significant. By incorporating all together, a better win rate of 84.75% was obtained.

Only-One-Victor Pattern Learning in Computer Go

Automatically acquiring domain knowledge from professional game records, a kind of pattern learning, is an attractive and challenging issue in computer Go. This paper proposes a supervised learning method, by introducing a new generalized Bradley-Terry model, named Only-One-Victor, to learn patterns from game records. Basically, our algorithm applies the same idea with Elo rating algorithm, which considers each move in game records as a group of move patterns, and the selected move as the winner of a kind of competition among all groups on current board. However, being different from the generalized Bradley-Terry model for group competition used in Elo rating algorithm, Only-One-Victor model in our work simulates the process of making selection from a set of possible candidates by considering such process as a group of independent pairwise comparisons. We use a graph theory model to prove the correctness of Only-One-Victor model. In addition, we also apply the Minorization-Maximization (MM) to solve the optimization task. Therefore, our algorithm still enjoys many computational advantages of Elo rating algorithm, such as the scalability with high dimensional feature space. With the training set containing 115,832 moves and the same feature setting, the results of our experiments show that Only-One-Victor outperforms Elo rating, a well-known best supervised pattern learning method.

On the Efficiency of Executing Diverse Game Tree Search Applications in a Volunteer Computing Federation

A game tree search application can be implemented as the malleable parallel jobs that are adapted to various processor allocations. We establish an efficient desktop grid federation to enable the small to mid-sized organizations to perform large-scale game tree search tasks via resource sharing. Due to the uneven task scales of the organizations as well as the dynamic generation/pruning of game tree search tasks, the user credits of the desktop grids may fluctuate dramatically, leading an unstable resource allocation to the hosted applications. This paper shows that stable processor allocation leads to higher efficiency for the parallel tasks. A new brokering algorithm is developed that ensures both fairness and stable resource allocation.

Connect6 Programs on Mobile Devices

In this demonstration, we port our current Connect6 program, NCTU6, to mobile devices, such as iOS devices, Android devices. However, computing powers of mobile devices are also relatively weaker when compared with high-end machines. For example, the CPU speeds of mobile devices are about 1/5 to 1/8 of those of desktops, and the memory of mobile devices is usually limited to several megabytes. In this demonstration, we tune the program NCTU6 to fit in mobile devices, while maintaining reasonable strength. The program also won the silver medal in TCGA 2012. In addition, we also support Connect6 puzzles in the program.