Using Automated Replay Annotation for Case-Based Planning in Games

A major challenge in the field of case-based reasoning is building case libraries representative of the state space the system will encounter. We approach this problem by automating the process of converting expert demonstrations, in the form of game replays, into cases. To achieve this, we present a technique for annotating traces with goals that can be used by a case-based planner. We have implemented this technique in the case-based planning system Darmok 2 and applied it to the task of playing complete games of StarCraft. By automating the process of case generation, we enable our system to harness the large number of expert replays available on the web.