Adaptive task-driven game AI for real-time strategy games

Abstract

RTS games provide a constant stream of decision-making under incomplete information. The number of possible game states is many orders of magnitude larger than that of Chess or Go, which means that applying techniques used in traditional board games to RTS games will result in inefficient solutions unless the representation of the game is simplified to fit these techniques. The RTS environment is mirrored by multiple real-life scenarios such as rescue operations and road traffic. Therefore developing an AI capable of working under such conditions can benefit other fields as well. ACT-AI is a task-driven solution for StarCraft facilitated by the BWAPI and BWTA2 third party libraries. The AI provides an abstraction of in-game actions using tasks, and optimises a set of unit compositions to be used by coalitions accomplishing those tasks. Agents join coalitions if they can contribute to the task. Composition fitness depends on the cost and profit of a coalition. By re-watching a match through the built-in replay system, compositions are modified according to the enemy's current unit composition. ACT-AI is evaluated over 300 games in three matchups, with 100 games per match-up, and 10 games per composition iteration. Results show that ACT-AI performs well against the Blizzard Zerg AI and poorly against the Blizzard Terran AI. The early stages of the game are a clear point of vulnerability. Score ratios show that the losses sustained by ACT-AI are close losses. The resulting composition fitness does not directly reflect the win-rates and score ratios. This may be due to composition fitness being dependent on the previous composition iteration, therefore indicating a trend rather than a representation of the current state. This result also calls into question the suitability of the fitness function, an avenue for future work. ACT-AI has poor economy management, but excellent resource acquisition capabilities, pooling excessive amounts of resources throughout the game. The lack of complex unit control is a likely contributor to this problem, preventing the AI from spending as desired by blocking units from reaching their destination. A defense map is implemented in order to prioritize defensive locations. Results for the defense map show a successful interpretation of important map locations such as chokepoints and resource locations.