Computational Conceptual Change: An Explanation-Based Approach.

摘要

The process of conceptual change - whereby new knowledge is adopted in the presence of prior conflicting knowledge - is pervasive in human cognitive development, and contributes to our cognitive flexibility. At present, Artificial Intelligence systems lack the flexibility of human conceptual change. This is due in part to challenges in knowledge representation, belief revision abduction, and induction. In addition, there are disagreements in the cognitive science community regarding how people represent, use, and revise their mental models of the world. This work describes a cognitive model of conceptual change. The claims are that (1) qualitative models provide a consistent computational account of human mental models, (2) our psychologically plausible model of analogical generalization can learn these models from examples, and (3) conceptual change can be modeled by iteratively constructing explanations and using meta-level reasoning to select among competing explanations and revise domain knowledge. The claims are supported by a computational model of conceptual change, an implementation of our model on a cognitive architecture, and four simulations. We simulate conceptual change in the domains of astronomy, biology, and force dynamics where examples of psychological conceptual change have been empirically documented. Aside from demonstrating domain generality, the simulations provide evidence for the claims of the thesis. Our simulation that learns mental models from observation induces qualitative models of movement, pushing, and blocking from observations and performs similar to students in problem-solving. Our simulation that creates and revises explanations about the changing of the seasons shows that our system can assemble and transform mental models like students. Our simulation of textbook knowledge acquisition shows that our system can incrementally repair incorrect knowledge like students using self-explanation.