Machine Learning in Real-Time Strategy Games

Abstract

We set out to take a look at machine learning in RTS-games, with an emphasis on dynamic scripting, and suggest some additions or improvements to the methods used up until now, and give some direction to future research in this area. Several additions and changes to the dynamic scripting algorithm will also be discussed, including improving the knowledge base with the aid of an evolutionary algorithm and altering it into a case-based reasoning system. These and other methods will be discussed in-depth. We see that dynamic scripting on its own, although performing decently against balanced scripts, does not deal well with optimized scripts. We also see that the additions and alterations to the algorithm improve it’s performance against these types of opponents. We conclude to suggest a number of avenues for future research, including improvements to opponent modelling for RTS-games, and communication systems between allied dynamic scripts.