Video games often have challenging and exciting computer-controlled
opponents to play against a human player. However, creating a artificial
intelligence (AI) to operate these opponents requires significant
development effort by both programmers and game designers. One solution is
to use automated processes to evaluate and improve the strength of AIs, and
adapt them to changing game maps or mechanics.
This seminar will present ongoing work into the application of genetic
algorithms to the AI of Battle for Wesnoth, a popular turn-based strategy
game. AI strategies are represented by a tree with action nodes, based on
the behaviour tree formalism. These strategies were then evaluated in a
Battle for Wesnoth game, in order to obtain a metric of their performance.
The genetic algorithm then preferentially selected and combined strategies
that performed well, allowing elements from strong strategies to transfer
into other strategies. By iterating this process, this algorithm produced
strategies that are stronger than the starting strategies.
An overview of genetic algorithms will be provided at the start of the
seminar, followed by a short discussion on the difficulty of evaluating AI
performance. Preliminary results from this work will then be presented,
including the strength of automatically-generated strategies versus