Adaptive Artificial Intelligence in Real-Time Strategy Games
Jason M. Traish · Charles Sturt University Research Output (CRO) · 2017
Highly capable Artificial Intelligences (AI) have been created for board games such as Go and Chess. Players of these games can play against a computerised opponent at the equivalent skill level of grandmaster or better. However, such highly capable AI agents have not yet been developed for Real-Time Strategy (RTS) games. RTS agents must address several challenging issues to demonstrate real player like skill. The first major issue is that RTS games play out in ’real-time’. In the RTS game context, ’real-time’ means that games do not have a rigid turn-based structure but play continuously with players taking actions at any time. The sec- ond major issue is that RTS games have a much larger game state-space than Go or Chess. This is because typical RTS games occur on large maps with terrain differentiations, and involve a large number of diverse units. They also involve many more actions such as resource collection, production management and dif- ferent types of tactical actions. Finally, unlike Go or Chess, players in RTS games possess only incomplete game-state information. RTS AI is one of the next big AI challenges. This thesis seeks to improve the quality of adaptive tactical RTS agents, by en- abling them to respond more effectively to novel player actions. This is intended to improve both the challenge for skilled players and the value of the single- player experience. The main focus of the thesis is on the issue of real-time decision making and the ability to adapt to changes within a game. The thesis provides a framework that allows an RTS AI to adapt to unknown scenarios without perceptible lag.