Searching Over Search Trees for Human-AI Collaboration in Exploratory Problem Solving: A Case Study in Algebra

Artificial intelligence and machine learning work very well for solving problems in domains where the optimal solution can be characterized precisely or in terms of adequate training data. However, when humans perform problem solving, they do not necessarily know how to characterize an optimal solution. We propose a framework for human-AI collaboration that gives humans ultimate control of the results of a problem solving task while playing to the strengths of the AI by persisting an agent's search trees and allowing humans to explore and search this search tree. This allows the use of AI in exploratory problem solving contexts. We demonstrate this framework applied to algebraic problem solving, and show that it enables a unique mode of interaction with symbolic computer algebra through the automatic completion and correction of traditional derivations, both in digital ink and textual keyboard input.