Constructing and Refining Causal Explanations from an Inconsistent Domain Theory

Richard J. DoyleAAAI 1986

Abstract

This paper challenges the traditional assumption in machine learning that a perfectly consistent domain theory is required for effective explanation-based learning. It proposes a framework where causal explanations are constructed from simplified, layered abstractions and refined through a failure-driven process when real-world inconsistencies are encountered.

Introduction

In the mid-1980s, Explanation-Based Learning (EBL) was a dominant paradigm, relying on a system's existing knowledge (domain theory) to explain why a specific example fits a concept. However, EBL faced a significant hurdle: the requirement for a consistent domain theory—a complete and perfect rulebook of the world. Doyle (1986) argues that real-world knowledge is rarely consistent. Instead, human reasoning utilizes abstractions, defaults, and approximations to maintain tractability. This research addresses how an AI can construct useful causal models using these "inconsistent" simplified theories and "drill down" into more detailed layers of knowledge when failures occur.

Layered Abstractions of Causal Mechanisms

Rather than a flat knowledge base, the system organizes causal mechanisms into a hierarchy of increasingly detailed layers. This structure allows the system to suppress unnecessary detail until a specific task or failure requires it.

Figure 1: The hierarchy of causal mechanisms used to organize domain knowledge into layered abstractions.

The hierarchy includes levels such as:

  • Co-occurrence: Simple temporal sequencing of events.
  • Quantity Types: Modeling how changes in one amount cause changes in another.
  • Medium: Identifying physical links (e.g., pipes, wires) between cause and effect.
  • Barrier: Introducing elements that can block a link (e.g., valves, switches).
  • Variable Barrier: Modeling barriers whose effectiveness depends on specific quantities.

Failure-Driven Refinement

The system operates by starting with the most abstract model possible. If the model fails to solve a planning problem or makes an incorrect prediction, a refinement process is triggered:

  1. Reinstantiation: Adding a new instance of a known element (e.g., identifying an additional barrier like a lens cap).
  2. Elaboration: Moving to a deeper, more detailed layer of the theory (e.g., modeling the safety latch on a mechanical coupling).

Case Studies: The Bathtub and the Camera

The system's logic is demonstrated through qualitative examples involving a bathtub and a camera. These examples illustrate how the system generalizes mechanisms across domains. For instance, the system identifies that a "valve" is a functional role played by a physical object acting as a barrier to a flow.

Figure 3: A causal model of a bathtub where a plug acts as a barrier to the flow of water from the tub to the drain.

In the bathtub example, the system initially fails to account for the drain. By refining the model to include the plug as a barrier, it successfully generates a plan to fill the tub. This "valve" concept is then transferred to the camera domain, where a shutter performs the same causal role for light flow as the plug does for water flow.

Key Insights

  • Inconsistency as Abstraction: Doyle reframes inconsistency not as a logical error, but as a byproduct of necessary abstraction. Reasoning is faster when detail is suppressed, provided the system can "un-suppress" it upon failure.
  • Task-Relative Consistency: An explanation only needs to be consistent relative to the specific inferences required for a goal, not "true" in an absolute sense.
  • Cross-Domain Analogy: By recognizing universal patterns like the "valve," the system demonstrates a form of analogical reasoning that allows it to apply learned causal structures to entirely different physical domains.

Conclusion and Future Work

The paper demonstrates that layered domain theories allow for robust reasoning in the face of simplified knowledge. However, limitations remain, such as "bottoming out" when a theory is incomplete even at its most detailed level. Future work suggests integrating inductive learning or active experiment design to distinguish between multiple plausible explanations when failures occur.

Summary updated

BibTeX
@inproceedings{doyle1986constructing,
  author = {Richard J. Doyle},
  title = {{Constructing and Refining Causal Explanations from an Inconsistent Domain Theory}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1986},
}
Figure