Project Aims: Determine whether learning agents in Real-Time Strategy (RTS) games are more challenging, realistic and diverse than the more classical AI techniques used in industry

Project Methods: (1). Develop an simplified RTS game, write some hand-coded human-level strategies into some example agents and test them against one another.
(2). Construct some hand-coded representations of human gameplay strategies using centralized and localized control to determine the suitability of A-life as a design template
(3). Develop an agent which uses a combination of Reinforcment Learning and A-life, train & test it against all of the above opponents, as well as a human player.
(4). Evaluate the realism of these agents using criteria such as exploiting weaknesses, element of surprise and diversity or sporadity of overall gameplay

Project Results & Conclusions: The learning agents were not only vastly more realistic than the others, but were quick & easy to train and discovered some gameplay design bugs in the RTS testbed, and exploited these to its advantage.

Future Work: Test these agents on a full-scale RTS game and other game genres, as the framework is extremely general and can be applied to most other game types.
Use this framework to develop a tool for automating game-testing, given its observed performance in the RTS game used for testing.