Learning to play Starcraft with Case-based Reasoning: Investigating issues in large-scale case-based planning

Abstract

In this master thesis we describe our work in creating a planner for the real-time strategy game Starcraft using case-based reasoning. Our work has been focused on the challenges in creating a usable casebase, and the resulting issues arising from scaling up the casebase.First, we present an agent designed to play Starcraft using plans from our CBR planner, and its architecture. We then move on to describe how this planner works, and how it overcomes the challenges in scaling up.We then present several experiments designed to measure how well our approach works given the limitations we have set. Finally, we discuss our results, and provide some interesting unsolved challenges which may benefit from further investigation.