Combining expert knowledge and learning from demonstration in real-time strategy games

Abstract

Case-based planning (CBP) is usually considered a good solution to solve the knowledge acquisition problem that arises when developing AIs for real-time strategy games. Unlike more classical approaches, such as state machines or rule-based systems, CBP allows experts to train AIs directly from games recorded by expert players. Unfortunately, this simple approach has also some drawbacks, for example it is not easy to refine an existing case base to learn specific strategies when a long game session is needed to create a new trace. Furthermore, CBP may be too reactive to small changes in the game state and, at the same time, do not respond fast enough to important changes in the opponent’s strategy. We propose to alleviate these problems by letting experts to inject decision making knowledge into the system in the form of behavior trees, and we show promising results in some experiments using Starcraft.