Towards Chunking as a General Leaming Mechanism John E. Laird, Paul S. Rosenbloom, and Allen Newell, Carnegie Mellon University

AAAI 1984

TL;DR

This paper proposes "chunking" as a universal learning mechanism within the Soar problem-solving architecture. By automatically creating and applying "chunks" (learned production rules) for every subgoal, Soar demonstrates significant performance improvements, across-task transfer of learning, and even strategy acquisition, moving beyond simple speed-ups to more sophisticated learning.

What Problem Does This Paper Address?

The paper tackles the challenge of creating a truly general learning mechanism for artificial intelligence systems. While previous work has explored chunking for human memory and skill acquisition, the authors aim to extend it to a broader range of complex cognitive tasks. They question whether a single, task-independent learning mechanism can underpin all forms of learning, including strategy acquisition and knowledge compilation, within a unified problem-solving framework. The goal is to move beyond specialized learning modules and develop a system that can learn about all aspects of its own behavior.

Methods

The core of the paper is the integration of chunking into the Soar architecture. Soar is a production-system-based problem solver that views all activity as heuristic search in problem spaces. When Soar encounters a difficulty or lacks sufficient knowledge to proceed, it automatically creates a subgoal to resolve the issue.

The novel aspect here is how chunking is applied:

  • Universal Subgoaling: Soar's ability to create subgoals for any aspect of its problem-solving behavior (e.g., selecting an operator, applying an operator, testing for goal satisfaction) means that all of these activities are open to learning.
  • Single-Production Chunks: Unlike earlier models, Soar creates single production rules (chunks) for each terminated subgoal. These chunks capture the learned knowledge.
  • Implicit Generalization: Chunks are built based on the data accessed during subgoal processing. This means they implicitly generalize by only including conditions relevant to the outcome, ignoring irrelevant contextual details.
  • Strategy Acquisition: By chunking search-control decisions and operator applications, Soar can learn efficient strategies, effectively transforming search-based problem-solving into more direct methods.

Datasets

The paper does not introduce new datasets but rather uses established AI benchmark tasks for evaluation:

  • Eight Puzzle: A classic sliding tile puzzle used to demonstrate speed-ups and within-trial transfer.
  • Tic-Tac-Toe: Used to showcase learning of symmetries and strategy acquisition.
  • R1 (part of): A portion of the R1 expert system for computer configuration was implemented to test chunking's ability to support knowledge-intensive systems and demonstrate learning with minimal search control.

Key Insights and Ideas

  • Chunking as a Universal Mechanism: The paper strongly argues that chunking, when integrated with a universal subgoaling architecture like Soar, can serve as a single, powerful learning mechanism for a wide range of cognitive tasks, not just simple skill acquisition.
  • Learning About Learning: Soar doesn't just learn how to solve a task; it learns about its own problem-solving process by chunking subgoals related to search control, operator selection, and goal satisfaction. This is a crucial step towards meta-learning.
  • Implicit Generalization is Key: The implicit generalization of chunks, by focusing on accessed data, is what enables significant transfer of learning, both within a single task (within-trial) and across different tasks (across-task). This is particularly evident in Tic-Tac-Toe where symmetries are learned without explicit programming.
  • Avoiding the "Wandering Bottleneck": By opening all aspects of behavior to learning, Soar avoids the problem where fixing one performance bottleneck simply shifts the bottleneck elsewhere. Chunking can address bottlenecks across the entire system.
  • Over-generalization as a Trade-off: While implicit generalization is powerful, it can lead to over-generalization, where a chunk becomes too broad and produces incorrect results in slightly different contexts. This is identified as a significant challenge.

Key Results

  • Performance Speed-ups: Experiments showed dramatic reductions in the number of decisions (a proxy for time) required to solve tasks after chunking, especially in the Eight Puzzle and R1. For example, in R1, after chunking, a task that initially took 1731 decisions was solved in just 7 decisions on subsequent runs.
  • Within-Trial Transfer: Chunking enabled learning to occur within a single task session, leading to faster performance on subsequent attempts within that same session.
  • Across-Task Transfer: Chunks learned in one task were shown to be applicable to a different task, demonstrating that chunking can capture generalizable knowledge.
  • Strategy Acquisition: In Tic-Tac-Toe, chunking led to the acquisition of strategic knowledge, allowing Soar to learn winning moves and exploit symmetries without explicit programming.

Limitations and Future Work

  • Over-generalization: The authors acknowledge that implicit generalization can lead to over-generalization, where chunks become too broad and produce incorrect results. They propose future work on recovering from over-general knowledge and learning discrimination rules, drawing parallels to Anderson's work on knowledge compilation.
  • Understanding Human Learning Assumptions: The paper notes that some assumptions of earlier chunking models for human learning were abandoned, and further research is needed to understand their necessity.
  • Characteristics of Problem Spaces: More investigation is needed into the properties of problem spaces that facilitate interesting forms of generalization, such as the use of symmetry.
  • Broader Learning Phenomena: The authors suggest that future work should cover other forms of learning not demonstrated in this paper, including concept formation, problem space creation, and learning by analogy, to fully establish chunking as a general learning mechanism.

Summary updated

BibTeX
@inproceedings{1984towards,
  title = {{Towards Chunking as a General Leaming Mechanism John E. Laird, Paul S. Rosenbloom, and Allen Newell, Carnegie Mellon University}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1984},
}

Loading related content...

Figure