Learning intermediate goals for human problem solving
Learning problem solving knowledge is a technique often used to improve efficiency of problem solvers in areas, such as behavior cloning, planning, or game playing. In this paper, we focus on learning problem solving knowledge that explains computer problem solving and can be used by a human to solve the same problems without a computer. We describe an algorithm for learning strategies, where a strategy is a sequence of subgoals. Each subgoal is a prerequisite for the next goal in the sequence, such that achieving one goal enables us to achieve the next goal with a limited amount of search. Strategies are learned from a state-space representation of the domain and a set of attributes used to define subgoals. We first demonstrate the algorithm on a simple domain of solving mathematical equations, where we use the complete state-space to learn strategies. In the other two domains, the 8-puzzle and Prolog programming, we introduce an extension of the algorithm that can learn from a subspace of states determined by example solutions.
Paper submitted to journal.
