Knowledge Level Learning in Soar

Paul S. Rosenbloom, John E. Laird, Allen NewellAAAI 1987

Introduction

In the late 1980s, a fundamental debate permeated the field of artificial intelligence: could a system truly learn something "new," or was learning merely a process of optimizing existing knowledge? This distinction, formalized by Thomas Dietterich, separated Symbol Level Learning (improving performance through speed-ups) from Knowledge Level Learning (acquiring facts not previously implied by the system's knowledge base).

At the center of this debate was Soar, an architecture proposed by Paul Rosenbloom, John Laird, and Allen Newell as a candidate for a "Unified Theory of Cognition." Soar utilized a mechanism known as chunking—a form of explanation-based generalization—to cache the results of its problem-solving efforts. While chunking was highly effective at making Soar faster, critics argued it was restricted to the symbol level. In their 1987 paper, "Knowledge Level Learning in Soar," the authors demonstrated that this architectural constraint was a misconception, showing how Soar could indeed acquire entirely new declarative knowledge.

The Mechanism: Impasses and Chunking

To understand how Soar learns, one must first understand how it acts. Soar operates by applying operators to states within "problem spaces." When the system lacks the knowledge to choose an operator or proceed with a task, it reaches an impasse. This impasse automatically triggers the creation of a subgoal to resolve the uncertainty.

Once the subgoal is resolved through search or further problem-solving, Soar’s chunking mechanism creates a new production rule (a "chunk"). This rule summarizes the conditions that led to the successful resolution of the subgoal. In future similar situations, the chunk fires, allowing the system to bypass the impasse and the subsequent search, thereby improving efficiency.

Data Chunking: Recognition and Recall

The core innovation of the paper is the introduction of Data Chunking. The authors framed declarative memory tasks—specifically recognition and recall—as problem-solving tasks within Soar's architecture.

Recognition

To learn to recognize a new object (for example, a specific sequence of letters), Soar is presented with the object's components. Because it does not yet have a single internal symbol for the whole structure, it hits an impasse. In the resulting subgoal, it "recognizes" the parts and their relationships, eventually assigning a unique internal name to the composite structure. The resulting chunk effectively says: If these specific parts are present in this configuration, then this is Object X. This creates a new "fact" in the system's long-term memory.

Recall

Recall is more complex. To "memorize" an object for later retrieval without it being present in the environment, Soar must be able to reconstruct it. The authors identified a significant hurdle: if a chunk is created while the target object is visible, the chunk's conditions will likely require the object to be visible to fire—rendering it useless for recall.

To solve this, they implemented a Generate-and-Test split. During the learning phase:

  1. Generate Phase: The system uses its known primitives to build a candidate object.
  2. Test Phase: It compares this candidate to the target object it is trying to memorize.
  3. The Chunk: Because the "test" phase only confirms the correctness of the "generate" phase's output, the resulting chunk captures the sequence of actions needed to build the object. Crucially, the chunk does not depend on the target object being present; it only depends on the intent to recall that specific identifier.

Breaking the Deductive Closure

The philosophical breakthrough of this work lies in how it handles the "deductive closure." Critics argued that because chunking is a deductive process, the resulting chunks are already "contained" within the system's initial rules.

Rosenbloom et al. argued that while the process of chunking is deductive, the inputs to the system—the perceptions of new objects—are not. By using problem-solving to map external perceptions to internal symbols, Soar captures information from the environment that was not previously part of its internal model. This allows the system to move beyond the symbol level and achieve genuine knowledge-level learning.

Experimental Results and Limitations

The authors validated their approach using hierarchical symbolic objects (e.g., nested lists of letters). They demonstrated that Soar could:

  • Learn to recognize complex, multi-level structures.
  • Recall these structures on demand.
  • Improve its processing speed for known structures (symbol-level gain) while simultaneously acquiring the ability to handle previously unknown structures (knowledge-level gain).

However, the paper also noted several limitations. The early implementation of recall was computationally expensive, as the system would essentially "re-generate" every object it knew when prompted to recall. Furthermore, the "perception" of objects was simulated, and the system's ability to perform more complex inductive generalizations (like moving from specific instances to general categories) remained a subject for future research.

Conclusion

"Knowledge Level Learning in Soar" remains a landmark paper in cognitive architecture. It proved that a single, relatively simple learning mechanism—chunking—could serve dual purposes: optimizing performance and acquiring new knowledge. By framing "remembering" as a form of "problem-solving," Rosenbloom, Laird, and Newell provided a powerful template for how autonomous agents can learn from their experiences in the world.

Summary updated

BibTeX
@inproceedings{rosenbloom1987knowledge,
  author = {Paul S. Rosenbloom and John E. Laird and Allen Newell},
  title = {{Knowledge Level Learning in Soar}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1987},
}
Figure