We are interested in the issues of learning in the context of problem solving. Analogy and reflection are the two main strands of our work on learning. Our work on the IDEAL project has led to a model-based approach to analogical reasoning and learning. Similarly, our work on reflection in the AUTOGNOSTIC project has led to a model-based approach to task-directed failure-driven learning.
Mental models are schemes for representing and organizing knowledge and analogy is a reasoning strategy. Both play central roles in reasoning and learning. The IDEAL project explores interactions between device models and analogy in the context of innovative design. IDEAL is a computational theory that provides an account of both how models enable analogical transfer and how analogies enable the construction of models. It uses two kinds of models: structure-behavior-function models of specific devices, and behavior-function models of generic physical processes and engineering mechanisms. The generic BF models are learned by abstraction from device-specific SBF models and mediate cross-domain analogical transfer. SBF models of new devices are acquired in part by adapting the models of similar devices and partly through analogical transfer in the form of BF models. Current work focuses on using IDEAL's computational theory of model-based analogy for understanding specific examples of scientific discovery in which both mental models and analogy apparently played a critical role.
The design of self-organizing systems capable of learning from their experiences and appropriately reorganizing their world knowledge and reasoning strategies is a central goal of AI. The AUTOGNOSTIC project explores the issues of functional and strategic self-redesign in AI systems. The key idea is that AI systems, such as ROUTER, can be viewed as abstract devices and failure-driven learning can be viewed as device redesign problem solving. And, in analogy to model-based design adaptation and device redesign in KRITIK, a meta-model of how the problem solving in the AI system works can enable the system to reflect on its experiences, especially its failures, and to redesign the problem solver. AUTOGNOSTIC is a computational model of model-based self-redesign. It provides a language for representing the structure and functions of a problem solver, and also its internal behaviors that specify how the functions of the problem solver are composed out of the functions of its structural elements. Experiments with AUTOGNOSTIC-on-ROUTER show that this scheme enables the system to redesign its organization and processing so as to incrementally improve its performance and enhance the range of problems it can solve. Current work focuses on identifying additional redesign and learning strategies.
For links to a complete list of papers and some ftp'able versions of the above papers go to: Complete Listing