AFRL-AFOSR-JP-TR-2016-0049 Understanding how to build long-lived learning collaborators

This project conducted basic research aimed at creating software systems that can collaborate naturally with people over extended periods of time. This involved investigating how to make a habitable combination of natural language and sketch understanding that supports interactive learning of complex domains, including giving advice, learning by reading, and learning by demonstration. We developed the notion of type-level qualitative representations that significantly improve expressive power and compactness, both of which improve reasoning and learning, while also providing a simpler path for learning qualitative models from natural language. We also made progress on using qualitative representations for strategic thinking, where continuous processes and causal knowledge about quantities provide a higher level of description, within which specific planning goals arise. This includes expressing goals in terms of maximizing/minimizing quantities, recognizing and analyzing tradeoffs, and encoding broader-scale strategies in terms of continuous processes. We explored how to extend the Companion cognitive architecture to incorporate more self-learning, including automatic detection of near-miss examples to improve discrimination in learning, and dynamic encoding strategies to improve visual encoding for learning via analogical generalization. We showed that spatial concepts can be learned via analogical generalization. Moreover, we explored learning sketched concepts via analogy at a larger scale than ever before, using a 20,000 sketch corpus to examine the tradeoffs involved in visual representation and analogical generalization. DISTRIBUTION A: Distribution approved for public release.