Exploring Game Balance and Tactics with AI in the Educational Wargame Counter-Air

Abstract: This report describes a study of the game balance of the wargame Counter-Air,using the artificial intelligence model AlphaZero. The project also analyzed the most effectivestrategies in the game regarding allocation of pieces to a set number of roles. Counter-Air is a board game being developed for use in tactical training of military officers. Inthe game, two players compete to destroy the most enemy pieces. Every round begins withallocation of remaining pieces to a set number of roles representing real-world aerial combat.The players then take turns firing upon each other. The game was modeled in OpenSpieland its implementation of AlphaZero was used to master the game to study the winfrequency and the average margin of victory, as well as the allocation tactics. The game was found to be slightly biased in favor of the attacking (blue) player, who won ona tiebreaker in games where the algorithm was further along in its training. Regardingallocation of pieces on the board, no major changes in strategy was observed as moretraining data was obtained. Allocation of pieces was somewhat distributed among the rolesfor both players, with the exception of blues Low Strike role, which was left completelyempty. Although the results indicate that AlphaZero has converged on an optimal solution ofthe game, analysis of the artificial neural network, and simulations in OpenSpiel wouldfurther ensure convergence. The tactics utilized by AlphaZero are for the most part realistic,however, they have not yet been analyzed by military professionals.