If you have any problems related to the accessibility of any content (or if you want to request that a specific publication be accessible), please contact (firstname.lastname@example.org). We will work to respond to each request in as timely a manner as possible.
Evolving Potential Fields to Direct Tactics in Real Time Strategy Games
AuthorOberberger, Michael Martin
Computer Science and Engineering
StatisticsView Usage Statistics
This thesis investigates the use of co-evolution to generate tactics in real-time strategy games. Games like Chess have been used to test AI approaches since the 1960s. Modern video games with simulated worlds now allow us to investigate AI approaches in less abstract spaces, thus allowing research results to be perhaps more immediately applicable. Real-time strategy games, a genre of modern video games, are widely used in wargaming and what-if scenario analysis in the military and in industry. Since good tactics can determine whether you win or lose, this thesis focuses on competent tactics generation in real-time strategy games.There are many ways to generate players for games. Many classical approaches employ a system of logic that relies on expert knowledge about the game and perhaps known effective strategies. Although they are good for certain kinds of problems, expert systems' approaches have been shown to be brittle and generally do not learn from experience. This work uses an evolutionary approach to learning tactics for RTS games. Since evolutionary techniques are generally good at learning solutions to specific problems, this work also employs co-evolutionary techniques to generate more robust tactics that are effective against potentially unseen opponents. My results from generating tactics against a specific opponent indicate that an evolutionary algorithm can evolve good tactics. The generated tactics defeat a known opponent after a relatively short training cycle. However, these tactics are specific to the opponents that were trained against - they do not perform as well against other opponents. Co-evolution leads to more adaptive tactics. My results from employing a co-evolutionary approach indicate that co-evolution can generate tactics that perform better over a set of previously unseen opponents. These results indicate the potential for co-evolutionary and evolutionary approaches to tactic generation. Because they may not be biased by human preconception, I believe that such approaches also have the potential to generate completely new and surprising tactical solutions to difficult problems.