Interactive Task Learning: A Cognitive Architecture Approach | Interdisciplinary Programs
Oct 26, 2017
4:10 PM - 5:30 PM
Packard Laboratory - 466
19 Memorial Drive West
Bethlehem PA 18015
Data X and Cognitive Science Program present Interactive Task Learning: A Cognitive Architecture Approach. John Laird the John L. Tishman Professor of Engineering, University of Michigan
Advances in AI are leading us to a future populated with intelligent agents that have the cognitive capabilities to perform intellectually challenging tasks in health care, business, the military, and the home. However, today’s agents are prisoners to the tasks for which they are programmed. Even though Deep Blue became the world’s chess champion and Watson won the Jeopardy Challenge, neither of them can be taught a new task, even something as simple as Tic-Tac-Toe. In contrast, humans quickly and continuously learn new tasks throughout our lifetimes.
In this talk, I describe progress in Interactive Task Learning (ITL), where an agent learns a novel task through natural interaction with an instructor. ITL is challenging because it requires a tight integration of many of the cognitive capabilities embodied in human-level intelligence: multiple types of reasoning, problem solving, and learning; multiple forms of knowledge representations; natural language interaction; dialog management; and interaction with an external environment – all in real time. Moreover, any successful approach must be general – the agent cannot be pre-engineered with the knowledge for a given task – everything about a task has to be learned or transferred from other tasks.
Our approach builds on our research with the Soar cognitive architecture. Soar provides the fixed, task independent computational infrastructure to support the integration of cognitive capabilities underlying human-level intelligence. Our approach to ITL emphasizes mixed initiative interaction, where the human provides advice and information, and the agent actively asks questions to acquire the knowledge it needs. I will describe how our agent (Rosie) processes instructions, learns internal representations of tasks, interprets those representations to perform the tasks, and dynamically compiles its knowledge for fast execution. I will show demonstrations of Rosie learning and solving puzzles and games (Tower of Hanoi, Eight Puzzle, Othello) from scratch for a table top robot as well as learning navigation tasks (deliver and fetch) for a small mobile robot, transferring its learned knowledge to new tasks. I’ll conclude with a discussion of the role ITL can have in the future of AI and some of the challenges it raises.
Please click "Additional Info" below for more information on this event.