|Speaker:||Dr. Dustin Dannenhauer, Naval Research Lab|
|When:||October 11, 2018, 11:00 am - 12:00 pm|
|Where:||Engineering Building, Room 3507|
Robust autonomy remains a difficult problem for complex, partially observable, and real-time environments such as robotics (e.g. mars rovers, underwater vehicles) and complex video games including Starcraft and Minecraft. Goal reasoning is a growing research area for controlling agents operating in these kinds of environments. Goal reasoning agents reason over their goals in response to unexpected changes in the environment. By monitoring their own execution, agents can detect if a situation is unexpected and if it will prevent them from achieving their goals. My research introduced new approaches for detecting unexpected situations (including failures) in the environment and in an agent's own cognitive processes via monitoring their own cognition (i.e., meta-cognition). In this talk I will give a brief background on goal reasoning and automated planning, walk-through an example of a construction agent in the Minecraft video game, highlight my work on anomaly detection for plan execution in goal driven autonomy agents, and conclude with exciting areas for future work.
Dr. Dustin Dannenhauer is a National Research Council Post-doctoral Researcher at the Naval Research Laboratory in Washington, D.C. He received his Ph.D. in Computer Science from Lehigh University in 2017 and his B.S. in Computer Science from Indiana University Bloomington in 2012. His research investigates autonomous agents that monitor their own execution for unexpected behavior. His work includes how semantically inferred concepts can be used as expectations in large, complex domains like real-time strategy games; how expectations can be generated and used to monitor plans produced from a hierarchical task network planner; and how expectations at the metacognitive level can be used to identify failures in cognition. Dr. Dannenhauer has been one of the core developers of the Metacognitive Integrated Dual-Cycle Architecture (MIDCA): a cognitive architecture with explicit metacognitive capabilities. He has published in top venues for Artificial Intelligence research including IJCAI, AAAI, ICCBR and Advances in Cognitive Systems.