Ontological Knowledge for Goal-Driven Autonomy Agents in Starcraft

Starcraft, a commercial Real-Time Strategy (RTS) game that has enjoyed world-wide popularity (including televised professional matches), is a challenging domain for automated computer agents. Evidence of this difficulty comes not only from characteristics of the game (massive state space, stochastic actions, partial visibility, etc.) but also from three years of competitive entries in tournaments (i.e. AIIDE Annual Starcraft Competition) in which the best automated entry performs poorly against a human expert. We are interested in taking a new research direction: using semantic knowledge, such as description logic, to represent the game state with abstract concepts in order to perform high level actions. Goal-driven autonomy (GDA) is a conceptual model for reasoning over goals in autonomous agents. The model makes use of a planner which produces both a plan and corresponding expectations. Expectations are what the agent expects to be true after executing actions in the plan. GDA reasons about goals in four components: discrepancy detection, explanation, goal formulation, and goal management. The process of a GDA agent begins with the goal management component sending a goal to the planner. The planner then produces a plan and expectations for what should be true in the future, as plan actions are executed. The agent begins executing the plan and, concurrently, the discrepancy detection component observes any expectations that are not satisfied. Upon identifying an unmet expectation, the explanation component generates an explanation for why the discrepancy occurred and sends the explanation to the goal formulation component, which may create a new goal. This new goal is sent to the goal management component, which then adjusts the priority of each goal and decides which goal(s) to send to the planner, and the cycle continues. An important aspect of the GDA model is that each component is open to implementation. Almost all previous work on GDA has been nonhierarchical (for example, discrepancy detection in the ARTUE (Molineaux, Klenk, and Aha 2010) system uses setdifference). Nor has previous work made use of an ontology. We propose using an ontology to represent the state of the