JuKeCB: Creating, Maintaining and Optimizing a Case Based Reasoning System

This paper discusses the research and creation of the case based reasoning system JuKeCB, a system which utilizes observation as a mean to create and store stochastic policies, delving into both its strengths and its weaknesses. The system was implemented into a domination style game where teams use various strategies to hold key map locations. By observing teams playing this game, the system is capable of learning which strategies are effective against others and reuses them appropriately when playing. Several extensions to the system are also discussed. These extensions use clustering and parallelization techniques and were designed to decrease the computation time of case retrieval---one of the systems main weaknesses. Several experiments were performed to extensively test both the system's performance against other teams and the speed at which it runs. The results of these experiments show that JuKeCB is capable of defeating many static policy opponents as well as other sophisticated AI agents with varying levels of training. Additionally they show that the extensions of JuKeCB increase the speed at which it can run without drastically impacting its performance.