A Model of Plan Inference that Distinguishes between the Beliefs of Actors and observers

Martha E. PollackACL 1986 · DOI

Introduction

In the mid-1980s, the field of Artificial Intelligence grappled with how to make computer systems more cooperative. A central technique was Plan Inference (PI), where a system attempts to deduce a user's underlying goals from their queries to provide more relevant assistance. However, early PI models suffered from a significant limitation: they assumed a "shared belief" model, where the user and the system possessed identical knowledge of the world. Martha Pollack’s seminal work challenged this, proposing a model that explicitly distinguishes between the beliefs of the actor (the user) and the observer (the system).

The Shift to a Mental Phenomenon View

Pollack argues that a plan is not merely a data structure—a sequence of actions to be executed—but a complex "mental phenomenon." Under this view, for an agent to have a plan, they must hold a specific configuration of beliefs and intentions. This shift allows a system to reason about why a user might propose an invalid plan. If a user asks a question based on a misconception, a system using Pollack's model can identify the discrepancy between the user's beliefs and the system's "ground truth."

Formalizing Intentions and Beliefs

The model utilizes formal logic to represent these mental states. Key predicates include:

  • BEL(A, P): Agent A believes proposition P.
  • INT(A, α): Agent A intends to perform action α.
  • EXEC(α): Action α is executable in the current state.
  • GEN(α, β): Performing action α "generates" (results in) action β.

By using these components, Pollack introduces Eplans (Explanatory Plans). An Eplan represents the system's hypothesis about the user's mental state, including the "explanatory beliefs" that lead the user to think their proposed actions will achieve their goal. This involves non-monotonic reasoning (via rules PI1 and PI2) to find "near-misses" in the system's knowledge base that explain the user's likely errors.

Diagnosing Plan Failure

A major contribution of this work is the categorization of how plans fail. When a system cannot reconcile a user's query with its own knowledge base, it classifies the failure into one of three types:

  1. Unexecutable: The user intends to perform an action that is impossible within the system (e.g., using a non-existent command).
  2. Ill-formed: The user can perform the actions, but those actions will not actually lead to the desired goal because the user's belief about the relationship between actions (the generation relation) is incorrect.
  3. Incoherent: The system cannot find any logical path, even accounting for potential misconceptions, that explains why the user would think their actions lead to their goal.

To handle these nuances, the model uses Conditional Generation (CGEN). This allows the system to represent that an action usually results in a goal, but fails to do so in the user's specific case because a necessary condition is false.

Implementation: The SPIRIT System

To demonstrate the model, Pollack developed SPIRIT (System for Plan Inference that Reasons about Invalidities Too). Operating in the domain of computer mail, SPIRIT could handle complex scenarios where users had incorrect mental models of system security.

For instance, if a user asked how to set permissions to block a system manager, SPIRIT would not just give a literal answer or a "command not found" error. Instead, it could generate a nuanced response: "That command won't keep Tom out: file permissions don't apply to the system manager. What you can do instead is move the file to a password-protected subdirectory." By identifying the user's false belief, the system provides the "missing link" of information required to solve the underlying problem.

Limitations and Legacy

Pollack’s work laid the groundwork for more sophisticated, user-aware AI. By acknowledging that users can be wrong, systems become significantly more helpful. However, the model does have limitations:

  • Observer Infallibility: The model assumes the system's knowledge is always the absolute truth.
  • Salience: The system lacks a mechanism to filter information based on what is most relevant to the conversation, sometimes providing more detail than a human would.
  • Complexity: Scaling this to open-ended domains requires more advanced reasoning about how beliefs evolve over time.

Despite these constraints, the distinction between actor and observer remains a cornerstone of cooperative interface design and human-computer interaction.

Summary updated

BibTeX
@inproceedings{pollack1986model,
  author = {Martha E. Pollack},
  title = {{A Model of Plan Inference that Distinguishes between the Beliefs of Actors and observers}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1986},
  doi = {10.3115/981131.981160},
}
Figure