Learning from Highly Flexible Tutorial Instruction

Scott B. Huffman, John E. LairdAAAI 1994

Introduction

Early "Learning Apprentice Systems" (LAS) were often limited by their inability to process instructions that deviated from direct, real-time expert demonstrations. This paper identifies three primary gaps in such systems: command rigidity, situation rigidity, and knowledge-type rigidity. To address these, the authors propose Instructo-Soar, an intelligent agent capable of learning from flexible, natural language tutorial dialogues through a process called situated explanation.

Taxonomy of instructional flexibility across command, situation, and knowledge-type dimensions.

Methodology: Situated Explanation

Instructo-Soar is built on the Soar architecture and utilizes the Problem Space Computational Model (PSCM). The core of its learning mechanism is "Forward Internal Projection," where the agent simulates instructed actions within its internal model to verify their progress toward a goal. This analytic approach, rooted in Explanation-Based Learning (EBL), allows the agent to ground instructions in both current and hypothetical states.

Handling Knowledge Gaps

A significant contribution of this work is how the agent manages incomplete explanations caused by missing knowledge ($M_K$). Rather than failing, Instructo-Soar evaluates the instructional context to select an appropriate resolution strategy.

Strategies for resolving incomplete explanations when the agent encounters missing knowledge (Mk).

As shown in the figure above, the agent has four primary options to resolve gaps:

  1. Delay: Postponing the explanation until the missing information is learned later.
  2. Induce: Using heuristics to make an educated guess.
  3. Ask: Requesting specific instruction from the teacher to fill the gap.
  4. Abandon: Falling back on simple inductive rules if a deep explanation cannot be constructed.

Evaluation and Results

The system was evaluated in a simulated robotic domain involving a "Hero" robot interacting with blocks, tables, and environmental controls. The agent demonstrated the ability to learn all five types of PSCM knowledge: state inference, operator proposal, preferential, effects, and termination.

Key results include:

  • Instructional Versatility: Instructo-Soar exhibited 19 distinct instructional capabilities and met 7 of 11 proposed requirements for a fully tutorable agent.
  • Efficiency: By leveraging EBL, the agent can learn general rules from a single example or instruction, avoiding the need for large datasets common in inductive learning.
  • Hypothetical Reasoning: The support for "what-if" scenarios (e.g., "If the light is off...") enables the agent to learn about rare or dangerous contingencies safely without physical execution.

Conclusion

Instructo-Soar represents a significant step toward creating agents that learn like human students. By applying situated explanation across varied contexts, the system achieves high flexibility in command handling, situational reasoning, and knowledge acquisition. Future work will focus on improving linguistic complexity and developing mechanisms for correcting erroneous prior knowledge.

Summary updated

BibTeX
@inproceedings{huffman1994learning,
  author = {Scott B. Huffman and John E. Laird},
  title = {{Learning from Highly Flexible Tutorial Instruction}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1994},
}
Figure