Creating large numbers of game AIs by learning behavior for cooperating units

Abstract

We present two improvements to the hybrid learning method for the shout-ahead architecture for units in the game Battle for Wesnoth. The shout-ahead architecture allows for units to perform decision making in two stages, first determining an action without knowledge of the intentions of other units, then, after communicating the intended action and likewise receiving the intentions of the other units, taking these intentions into account for the final decision on the next action. The decision making uses two rule sets and reinforcement learning is used to learn rule weights (that influence decision making), while evolutionary learning is used to evolve good rule sets. Our improvements add knowledge about terrain to the learning and also evaluate unit behaviors on several scenario maps to learn more general rules. The use of terrain knowledge resulted in improvements in the win percentage of evolved teams between 3 and 14 percentage points for different maps, while using several maps to learn from resulted in nearly similar win percentages on maps not learned from as on the maps learned from.