Learning a Game Strategy Using Pattern-Weights and Self-Play

Ari Shapiro UCLA ashapiro@cs.ucla.edu http://www.cs.ucla.edu/~ashapiro	Gil Fuchs UCSC	Robert Levinson UCSC levinson@cse.ucsc.edu

Abstract Automated computer players have been designed for many games, including chess, backgammon, Go and card games such as poker and bridge. For games such as chess and backgammon, the best computer programs can play at the level of the best human players. However, many games that have a large search space cannot be solved through search-based means and must use knowledge-based and feature-based approaches. We develop a feature-based automated player for the game of Diplomacy and demonstrate that features can be used to automatically discover plausible piece movements and game playing.
Download the paper
Learning a Game Strategy Using Pattern-Weights and Self-Play, Computers and Games, 2002
We have developed a Java-based online version of the game Diplomacy, called FreeDip and located on SourceForge.net. The software handles email contact, automatic turn resolution, graphical interface, and is capable of self-playing approximately 1000 games/day. The opening move of Diplomacy has over 4 trillion move combinations, making brute-force search of such space in a reasonable time unrealistic. Instead, the system decides moves based on game features called pattern-weights. Each feature is assigned a weighting, and combinations of features are used to decide upon the best course of action. For example, a feature such as: "unit is adjacent to a supply center" "supply center is unoccupied" would, after many games of self-play, eventually assign a high weighting to such a combination. This would bias the computer to choose a move into such a location. Both individual feature weightings and combinations are used to determine the best course of action. We conclude by showing that our pattern-weight player compares favorably with an human expert 'opening book' set of moves.