Using analogy to overcome brittleness in AI systems

One of the most important aspects of human reasoning is our ability to robustly adapt to new situations, tasks, and domains. Current AI systems exhibit brittleness when faced with new situations and domains. This work explores how structure mapping models of analogical processing allow for the robust reuse of domain knowledge. This work focuses on two analogical methods to reuse existing knowledge in novel situations and domains.;The first method, analogical model formulation, applies analogy to the task of model formulation. Model formulation is the process of moving from a scenario or system description to a formal vocabulary of abstractions and causal models that can be used effectively for problem-solving. Analogical model formulation uses prior examples to determine which abstractions, assumptions, quantities, equations, and causal models are applicable in new situations within the same domain. By employing examples, the range of an analogical model formulation system is extendable by adding additional example-specific models. The robustness of this method for reasoning and learning is evaluated in a series of experiments in two domains, everyday physical reasoning with sketches and textbook physics problem-solving.;The second method, domain transfer via analogy, is a task-level model of cross-domain analogical learning. DTA helps overcome brittleness by allowing abstract domain knowledge, in this case equation schemas and control knowledge, to be transferred to new domains. DTA learns a domain mapping between the entities and relations of the new domain and the understood domain, through comparisons between structures of explanations. Then, using this mapping, a new domain theory can be inferred and extended through an analogy between the domain theories themselves. This model is evaluated across a variety of physics domains (e.g., mechanics, electricity and heat flow). Successful cross-domain analogies result in persistent mappings, which support incremental learning of the target domain theory through multiple cross-domain analogies.