Towards Software Apprentices that Learn in Dynamic Domains


Sponsor: Computational Cognition and Machine Intelligence Program, Air Force Office of Scientific Research

Principal Investigator: Kenneth D. Forbus and Tom Hinrichs

Project Summary: T=Our goal is to discover how to create software apprentices, building on our work on the Companion cognitive architecture. Creating such intelligent systems that are enough like us to be apprentices, but different enough from us to provide complementary strengths, would be a highly disruptive breakthrough. Our specific objectives are (1) develop an account of the representations needed to support flexible learning, reasoning, and communication about strategies, tactics, and decision-making in dynamic environments; (2) Investigate the language understanding and sketch understanding capabilities needed to enable intelligent systems to interact with people as apprentices do, learning from demonstration, advice, and stories, without exposure to the internal representations of the system; and (3) Investigate how to build more autonomous intelligent systems, capable of managing their own learning and model-building capabilities, over a month of operation. Based on evidence from psychology and from our own prior computational investigations, our hypothesis is that analogical processing is central to human cognition, and thus should play a central role in intelligent systems. Our models of analogical matching, retrieval, and generalization provide human-like capabilities, which we will use with our work on qualitative representations, which provides a level of representation that bridges perception and cognition, supplies an important aspect of natural language semantics, and promotes learning general, transferrable knowledge. We will test these ideas by using both a strategy game, because it incorporates analogs of military, economic, and political phenomena, with complex temporal and spatial dynamics, as well as with multimodal stories, consisting of simplified English text and sketches. Being able to apply lessons learned in stories, and lessons learned in the game to other problems expressed via stories, will provide a strong test of the generality and utility of the accounts we develop.

Selected publications:

  1. Forbus, K. (2019). Qualitative Representations: How People Reason and Learn about the Continuous World, MIT Press.
  2. Nakos, C., Rabkina, I., Hill, S., & Forbus, K. (2020) Corrective Processes in Modeling Reference Resolution. In Proceedings of CogSci 2020, Online.
  3. Forbus, K., & Hinrichs, T. (2019). Qualitative Reasoning about Investment Decisions. Proceedings of the 32nd International Workshop on Qualitative Reasoning (QR 2019). Macao, China.
  4. Hinrichs, T. and Forbus, K. (2019). Experimentation in a Model-Based Game. In Proceedings of the Seventh Annual Conference on Advances in Cognitive Systems. Cambridge, MA.
  5. Forbus, K. & Hinrichs, T. (2018) Qualitative Reasoning for Decision-Making: A Preliminary Report. Proceedings of QR 2018.
  6. Forbus, K.D. & Hinrichs, T. (2017) Analogy and Qualitative Representations in the Companion Cognitive Architecture. AI Magazine.


Back to Projects page | Back to QRG Home Page