Towards Long-Lived Learning Software Collaborators
Sponsor: Robust Computational Intelligence Program, Air Force Office of Scientific Research
Principal Investigator: Kenneth D. Forbus and Tom Hinrichs
Project Summary: A key problem in achieving robust computational intelligent systems is conceptual complexity: Working collaboratively with people, scaling up reasoning and learning to handle the complexity and open-endedness of real-world phenomena, dealing with adversaries, and handling uncertain, incomplete, and buggy knowledge. Our goal is to move beyond the model of software as tool to the model of software as collaborator. The problems we face are getting harder, and people aren't getting any smarter. Creating intelligent systems that are enough like us to be effective, trusted collaborators, but different enough from us to provide complementary strengths, would constitute a highly disruptive breakthrough.
This project focuses on two bottlenecks:
- Interactive learning. Many important aspects of human learning occur via apprenticeship. Robust computational intelligences need to be able to participate in apprenticeship relationships, both cooperative and competitive, taking on increasing responsibility as they improve.
- Longevity. Today's AI and machine learning software are like drag racers, achieving high performance only along a single narrow dimension and requiring constant tinkering with their internals by experts to maintain them. Robust computational intelligences must be able to learn and adapt over substantial periods of time, improving their own knowledge and skills without the intervention of experts who understand their internals.
Our approach is to extend our Companion cognitive architecture, whose goal is to create software social organisms that interact with, and learn from, human partners. Our objectives are to scale up our analogical processing techniques to cumulatively learn substantial bodies of knowledge, investigate how to make a habitable combination of natural language and sketch understanding to support apprenticeship, and self-learning techniques to enable Companions to adapt and optimize their own performance in the face of changing workloads and disruptive environments over long periods of time.
- Hinrichs, R. & Forbus, K. (2012). Learning Qualitative Models by Demonstration. Proceedings of AAAI-12.
- Hinrichs, R. & Forbus, K. (2012). Toward Higher-Order Qualitative Representations.Proceedings of the 26th International Workshop on Qualitative Reasoning. Playa Vista, California.
- McLure, M. & Forbus, K. (2012). Encoding Strategies for Learning Geographical Concepts via Analogy. Proceedings of the 26th International Workshop on Qualitative Reasoning. Playa Vista, California.
- McLure, M. D., Friedman, S. E., Lovett, A. & Forbus, K. D. (2011). Edge-cycles: A qualitative sketch representation to support recognition. In Proceedings of the 25th International Workshop on Qualitative Reasoning. Barcelona, Spain.