Learning from Highly Flexible Tutorial Instruction

Scott B. Huffrnan and John E. Laird Artificial Intelligence Laboratory The University of Michigan Ann Arbor, Michigan 48109-2110 huffman@umich. edu instructions give 11 * Situated, interactive tutorial flexibility in teaching tasks, by. _ allowing communication of a variety of types of knowledge in a variety of situations. To exploit this flexibility, however, an instructable agent must be able to learn different types of knowledge from different instructional interactions. This paper presents an approach to learning from flexible tutorial instruction, called situated explanation, that takes advantage of constraints in different instructional contexts to guide the learning process. This makes it applicable to a wide range of instructional interactions. The theory is implemented in an agent called Instructo-Soar, that learns new tasks and other domain knowledge from natural language instructions. Instructo-Soar meets three key requirements of flexible instructability: it can (A) take any command at each instruction point, (B) handle instructions that apply to either the current situation or a hypothetical one (e. g. , conditionals), and (C) 1earn each type of knowledge it uses (derived from its underlying computational model) from instructions. at whatever point it is needed. The challenge is designing a tutorable agent that supports the wide breadth of interaction and learning abilities required by this flexible communication of knowledge. Our ultimate goal is to produce fully flexible tutorable agents, that can be instructed in the same ways as human students. A complete description of the properties and associated requirements of tutorial instruction is given by Huffman [1994]. In this paper, we focus specifically on three crucial requirements of flexible instructability: A. Command flexibility. The instructor should be free to give any appropriate commands to teach a task. Commands should not be limited to only directly performable/observable actions (as in [Redmond, 1992; Mitchell et al. , 1990; Segre, 1987]), but may request unknown procedures or actions that cannot be performed until after other actions.

Authors

• Scott B. Huffman John E. Laird

Keywords

No keywords are indexed for this paper.