Robust Learning and Reasoning for Intelligent Agents
Sponsor: Intelligent & Autonomous Systems Program, Office of Naval Research
Principal Investigator: Kenneth D. Forbus
Project Summary: We are exploring the use of qualitative modeling and analogical processing to provide robust common sense reasoning, learning, and communication capabilities for intelligent agents. Common sense reasoning is a necessary prerequisite to creating many kinds of useful intelligent agents that collaborate with human partners to accomplish tasks. Examples of such tasks include damage control assessment, operations planning, sifting through on-line information for relevant data, teaching and tutoring, and developing complex scientific, engineering, and policy models. Furthermore, intelligent systems need to communicate naturally with their human partners, interacting via natural language and sketching as appropriate. They must be able to learn, both when being brought up to speed in a new task and incrementally as a side-effect of working with people. They must take responsibility for their own learning to a large degree, since it is impractical to hand-optimize systems at the scale that will be needed for effective operation in many tasks. People who are domain experts should be able to serve as their teachers and trainers, without being AI experts. Intelligent systems should be able to take coaching for on-the-job learning from their collaborators.
We are focusing currently on learning by reading, i.e. being able to build up conceptual models of a domain that support broad, common-sense reasoning about it. Such reasoning goes beyond the kind of factoid Q/A which IBM's Watson performs. We are using our Companion cognitive architecture in conducting our experiments. The particular topics we are investigating are
- Knowledge-rich semantic interpretation. We are exploring techniques for automatically understanding article-length texts written in simplified English with sketches. This will include extending our query-driven abductive interpretation system with queries suitable for other genres of text, developing least-commitment representations for postponing disambiguation until more is known, and analogy-based techniques for recognizing occurrences of schema and generating evidential rules via generalization.
- Self-guided learning and optimization. We are developing techniques that enable systems to select a Sharma, A. and Forbus, K. (2012). Modeling the Evolution of Knowledge in Learning Systems. Proceedings of the Twenty-Sixth Conference on Artificial Intelligence (AAAI-12), Toronto, Canada.nd pursue their own learning goals, in order to improve their performance. This will include choosing new articles to read to acquire background knowledge, from the Simple English Wikipedia. It will also include learning new words to extend its reading abilities and improving its knowledge of quantities, to improve its estimation and error detection capabilities.
- Sharma, A. and Forbus, K. (2012). Modeling the Evolution of Knowledge in Learning Systems. Proceedings of the Twenty-Sixth Conference on Artificial Intelligence (AAAI-12), Toronto, Canada.
- Barbella, D. and Forbus, K. (2011). Analogical dialogue acts: Supporting learning by reading analogies in instructional texts. Proceedings of AAAI-11.
- Forbus, K., Hinrichs, T., de Kleer, J., and Usher, J. (2010). FIRE: Infrastructure for experience-based systems with common sense. Proceedings of AAAI Fall Symposium on Commonsense Knowledge, Arlington, VA.
- Sharma, A. and Forbus, K. (2010). Graph-based reasoning and reinforcement learning for improving Q/A performance in large knowledge-based systems. Proceedings of AAAI Fall Symposium on Commonsense Knowledge, Arlington, VA.
- Sharma, A. and Forbus, K. (2010). Modeling the evolution of knowledge and reasoning in learning systems. Proceedings of AAAI Fall Symposium on Commonsense Knowledge, Arlington, VA.
- Forbus, K., Lockwood, K. and Sharma, A. (2009). Steps towards a 2nd generation learning by reading system. AAAI Spring Symposium on Learning by Reading, Spring 2009.
- Lockwood, K. and Forbus, K. (2009). Multimodal knowledge capture from text and diagrams. Proceedings of KCAP-2009.