The Role of Prior Causal Theories in Generalization

Michael J. Pazzani, Michael G. Dyer, Margot FlowersAAAI 1986

Abstract: This classic 1986 research from the UCLA Artificial Intelligence Laboratory presents OCCAM, a system designed to unify two disparate paradigms in machine learning: correlational (similarity-based) learning and explanation-based learning. By leveraging prior causal knowledge and a set of heuristic generalization rules, OCCAM demonstrates how an intelligent agent can generalize effectively from a limited number of examples, bridging the gap between data-driven patterns and theory-driven understanding.

Introduction

In the mid-1980s, machine learning research was largely bifurcated. Similarity-based learners required large datasets to extract statistical correlations but were prone to noise and coincidental patterns. Conversely, explanation-based learners could generalize from a single instance but were brittle, requiring complete domain theories to function. OCCAM addresses this challenge by integrating both approaches, using existing knowledge to accelerate learning while maintaining the ability to "bootstrap" new theories in unfamiliar domains.

The OCCAM Architecture

OCCAM organizes memories of events into a hierarchical structure and follows a prioritized learning strategy when encountering new data:

  1. Explanatory Generalization: The system first attempts to explain an outcome using its existing causal theories. If a logical link is found, it creates a robust generalization immediately.
  2. Generalization Rules: If no prior theory exists, OCCAM applies heuristics—such as the DIFFERENT-FEATURES rule—to suggest potential causality based on temporal or physical cues.
  3. Tentative Generalizations: Hypotheses generated by rules are maintained as "working theories." These are held until they are either confirmed by subsequent data or refuted by contradictions.

To represent complex events like economic sanctions or interpersonal coercion, OCCAM utilizes structured symbolic frames that define the relationships between actors, objects, and goals.

Figure 1: A symbolic representation of a 'coerce' event schema in OCCAM, detailing the roles of actors, victims, threats, and the resulting goal failure due to third-party suppliers.

Experimental Domains

The system was evaluated across three distinct symbolic domains to test its flexibility:

  • Kidnapping/Coercion: Stories involving ransom demands were used to show how interpersonal relationships (e.g., a father's goal to protect a child) guide the learning of social schemas.
  • Physical World (Balloon Inflation): Based on protocols of a child attempting to inflate balloons, this domain demonstrated how OCCAM builds theories from scratch, eventually realizing that pre-stretched balloons are easier to inflate.
  • Economic Sanctions: Historical cases, such as the 1980 US grain embargo, were used to show how the system identifies relevant political features, such as the presence of a "third-party supplier" that causes sanctions to fail.

Key Insights and Results

  • The Causal Bias: Drawing on social psychology, the authors argue that humans are biased toward seeking causal explanations. OCCAM adopts this bias, allowing it to ignore irrelevant features (like the color of a victim's clothing) that would confuse a purely statistical learner.
  • Sample Efficiency: In the kidnapping domain, OCCAM could correctly generalize after seeing just one example by applying its interpersonal knowledge. A correlational learner would have required multiple examples to rule out coincidental traits.
  • Theory Bootstrapping: OCCAM successfully transitions from tentative hypotheses to sophisticated theories. In the economic domain, it filtered out irrelevant commonalities like "type of government" to focus on the economic mechanics of supply and demand.

Conclusion and Future Work

OCCAM represents a significant step toward human-like AI by prioritizing the search for "why" over the mere observation of "what." While the system was originally a passive observer, the authors noted that future iterations should incorporate active learning—such as suggesting experiments or querying authorities—to verify its tentative generalizations more efficiently.

Summary updated

BibTeX
@inproceedings{pazzani1986role,
  author = {Michael J. Pazzani and Michael G. Dyer and Margot Flowers},
  title = {{The Role of Prior Causal Theories in Generalization}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1986},
}
Figure