Dustin Dannenhauer

Toward Meta-Level Control of Autonomous Agents

Metareasoning is an important capability for autonomous systems, particularly for those being deployed on long duration missions. An agent with increased self-observation and the ability to control itself in response to changing environments will be more capable in achieving its goals. This is essential for long-duration missions where system designers will not be able to, theoretically or practically, predict all possible problems that the agent may encounter. In this paper we describe preliminary work that integrates the metacognitive architecture MIDCA with an autonomous TREX agent, creating a more self-observable and adaptive agent.

Case-Based Goal Selection Inspired by IBM's Watson

IBM’s Watson uses a variety of scoring algorithms to rank candidate answers for natural language questions. These scoring algorithms played a crucial role in Watson’s win against human champions in Jeopardy!. We show that this same technique can be implemented within a real-time strategy (RTS) game playing goal-driven autonomy (GDA) agent. Previous GDA agents in RTS games were forced to use very compact state representations. Watson’s scoring algorithms technique removes this restriction for goal selection, allowing the use of all information available in the game state. Unfortunately, there is a high knowledge engineering effort required to create new scoring algorithms. We alleviate this burden using case-based reasoning to approximate past observations of a scoring algorithm system. Our experiments in a real-time strategy game show that goal selection by the CBR system attains comparable in-game performance to a baseline scoring algorithm system.

Goal-Driven Autonomy with Semantically-Annotated Hierarchical Cases

We present LUiGi-H a goal-driven autonomy (GDA) agent. Like other GDA agents it introspectively reasons about its own expectations to formulate new goals. Unlike other GDA agents, LUiGi-H uses cases consisting of hierarchical plans and semantic annotations of the expectations of those plans. Expectations indicate conditions that must be true when parts of the plan are executed. Using an ontology, semantic annotations are defined via inferred facts enabling LUiGi-H to reason with GDA elements at different levels of abstraction. We compared LUiGi-H against an ablated version, LUiGi, that uses non-hierarchal cases. Both agents have access to the same base-level (i.e. non-hierarchical plans), while only LUiGi-H makes use of hierarchical plans. In our experiments, LUiGi-H outperforms LUiGi.