Analogical Learning for Companion Cognitive Systems
Funded by DARPA IPTO
We are creating Companion Cognitive Systems, interactive systems with human-like learning abilities that can serve as software aides-de-camp in a professional user’s work. Companions will help their users work through arguments and evidence (e.g., what an opponent’s strategy might be and how best to defeat them, an analyst understanding an international crisis, a scientist understanding a complex system), checking reasoning, looking for counter-evidence, and tracking dependencies as circumstances change over days, weeks, or months. Companions will learn knowledge and skills incrementally over time, increasing their understanding of the work and how better to interact with their user. This increasing shared awareness should result in a form of intelligence amplification. Companions will almost never be turned off, using non-interaction time to tackle longer assignments posed to them by their user, to reflect on and consolidate their knowledge, and to perform self-maintenance.
This project focuses on the core capability needed to create Companions: analogical learning. Our working hypothesis is that the flexibility and breadth of human common sense reasoning and learning arises from analogical reasoning and learning from experience. Within-domain analogies provide rapid, robust predictions. Analogies between domains can yield deep new insights and facilitate learning from instruction. First-principles reasoning emerges slowly, as generalizations created from examples incrementally via analogical comparisons. This hypothesis suggests a very different approach to building robust cognitive software than is typically proposed. Reasoning and learning by analogy are central, rather than exotic operations undertaken only rarely. Accumulating and refining examples becomes central to building systems that can learn and adapt.
Using our cognitive science research on analogy and similarity, we have created software components for analogical matching, similarity-based retrieval, and incremental generalization that have solid theoretical and empirical grounding. These components will form the core of the Companion architecture, which will perform analogical learning in three ways:
- Accumulation of examples. Companions will be able to
reason about new situations via analogies with prior examples. This will
provide a simple but effective way to learn even from single examples.
- Interactive analogical learning. Companions will be able
to learn from analogies presented by its human partner, to accelerate learning.
- Generalization from multiple examples. Companions will be able to refine new abstract knowledge from multiple examples, incrementally.
These learning mechanisms are being used for learning about the domain(s) the Companion operates in (e.g., strategy games and everyday physical reasoning), the preferences and work habits of its user, and how to improve its own performance. We are developing techniques for exploiting cluster computers to provide the combination of low-latency interaction, deep reasoning, and continuous operation over weeks and months that Companion systems will need.
Forbus, K. and Hinrichs, T. (2004). Self-modeling in Companion Cognitive Systems: Current Plans. DARPA Workshop on Self-Aware Systems, Washington, DC.
Forbus, K. and Hinrichs, T. (2004). Companion Cognitive Systems: A step towards human-level AI. To appear in AAAI Fall Symposium on Achieving Human-level Intelligence through Integrated Systems and Research, October, Washington, DC.