Why Things Go Wrong: A Formal Theory of Causal Reasoning

Leora Morgenstern, Lynn Andrea SteinAAAI 1988

Abstract

In the late 1980s, the field of Artificial Intelligence faced a significant hurdle in representing common-sense reasoning: the Frame Problem. While non-monotonic logics were designed to handle the assumption that things remain unchanged unless acted upon, they famously failed to resolve the Yale Shooting Problem (YSP). This paper by Morgenstern and Stein (1988) introduces a formal theory of causal reasoning based on the principle of "motivated actions." By preferring models where events only occur if they are triggered by known rules or observations, the authors provide a unified framework that successfully handles both forward prediction and backward explanation.

Introduction: The Frame Problem and the Yale Shooting Paradox

A central challenge in AI is the Frame Problem: the difficulty of logically representing the fact that most aspects of the world stay the same unless an action explicitly changes them. Early attempts to solve this using non-monotonic logic relied on the principle of inertia—the idea that properties persist over time. However, these systems were vulnerable to the Yale Shooting Problem (YSP).

In the classic YSP scenario, a gun is loaded at time $t_1$, a period of waiting occurs, and the gun is fired at Fred at time $t_5$. Intuitively, we expect Fred to be dead at $t_6$. However, standard logical frameworks of the era could not distinguish between the intuitive model and a "freak" model where the gun spontaneously becomes unloaded during the waiting period, allowing Fred to survive. Previous attempts to fix this, such as "chronological minimization," assumed that changes happen as late as possible. While this worked for forward prediction, it failed miserably at backward reasoning; if Fred were found alive at $t_6$, these theories would arbitrarily conclude the gun became unloaded at the very last possible moment, which does not align with human causal intuition.

The Theory of Motivated Actions

Morgenstern and Stein proposed a new preference criterion for models based on the concept of motivation. Their logic is built on three pillars:

  1. Causal Rules: These define how specific actions and preconditions lead to results (e.g., shooting a loaded gun results in a death).
  2. Persistence Rules: These explicitly state that properties persist unless an action occurs to change them.
  3. Motivation: This is the core innovation. An action is considered "motivated" if it is either required by the initial problem description (Strong Motivation) or triggered by a causal or persistence rule within a specific model (Weak Motivation).

Under this framework, the logic prefers models that contain the fewest possible unmotivated actions. Essentially, the system discounts any model where an event happens "just because." If an action occurs in a model without a clear cause defined in the rules or the initial facts, that model is considered less plausible.

Benchmarking Causal Logic

As a theoretical work in formal logic, the paper utilizes classic benchmarks from the "Commonsense Reasoning" community rather than empirical datasets:

  • The Yale Shooting Problem: Used to demonstrate that the theory correctly predicts Fred's death by discarding the unmotivated "spontaneous unloading" model.
  • Backward Projection (Modified YSP): A scenario where Fred is observed to be alive at $t_6$. The theory correctly infers that an "unload" action must have occurred, but it remains appropriately "ignorant" about the exact timing (it does not pick an arbitrary time like $t_4$),
  • Message Passing Scenario: A complex checkpoint-based problem used to show that minimizing unmotivated actions is more robust than simply minimizing the total number of actions.

Key Insights and Results

The primary contribution of this work is the shift from minimizing change to minimizing unmotivated change. This allows the theory to treat temporal reasoning as a two-way street. A robust theory of reasoning must be able to predict the future (forward projection) and explain the past (backward explanation) using the same underlying logic.

The authors define an "adequate explanation" as the difference between what was originally projected and what was actually observed. If a projection fails, the explanation is the set of facts—such as an unobserved "unload" action—that prunes away the incorrect models. This successfully links temporal projection with a formal theory of explanation, providing a rigorous definition for "why things go wrong."

Limitations and Future Directions

Despite its elegance, the 1988 theory has limitations. The persistence rules used in the framework are hand-coded; the authors noted that these should ideally be derived automatically from causal rules. Furthermore, the theory focuses on minimizing unmotivated actions. A more general approach might involve minimizing unmotivated state changes, which could potentially eliminate the need for the principle of inertia entirely. Finally, while the logic is theoretically sound, the paper does not address the computational complexity of finding preferred models in large-scale, real-world knowledge bases.

Summary updated

BibTeX
@inproceedings{morgenstern1988why,
  author = {Leora Morgenstern and Lynn Andrea Stein},
  title = {{Why Things Go Wrong: A Formal Theory of Causal Reasoning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1988},
}
Figure