Introduction
In the early development of artificial intelligence, planning systems and story-understanding programs relied heavily on domain-specific "world knowledge." These systems, such as Wilensky’s own Plan Applier Mechanism (PAM), functioned by mapping specific goals to known procedures—for example, knowing that to acquire money, one should go to a bank. However, these early models encountered significant friction when faced with multiple, interacting goals.
Standard planners often struggled to explain why an agent might deviate from a primary objective. Consider a scenario where a character, John, is speeding to reach a destination quickly but slows down upon spotting a police car. A rigid planner focused solely on the goal of "reaching the destination" cannot account for this change in behavior without a higher-level understanding of goal conflict—specifically, the interaction between the desire for speed and the desire to avoid a traffic citation. This 1980 paper introduces meta-planning, a theory proposing that the planning process itself must be treated as a domain for reasoning to handle such complexities.
The Meta-Planning Framework
Wilensky proposes that the same mechanisms used for ordinary planning can be applied recursively to the planning process. This framework is built upon two core components:
- Meta-goals: These are goals directed at the planning process rather than the external world. Examples include "Don't Waste Resources," "Achieve As Many Goals As Possible," and "Resolve Goal Conflicts."
- Meta-plans: These are strategies designed to satisfy meta-goals. For instance, if an agent identifies two distinct goals that both require travel, a meta-plan for "Plan Integration" might be triggered to combine the trips into a single, efficient route.
By representing planning knowledge declaratively as meta-goals and meta-plans, the system can detect "meta-situations"—specific patterns of goal interaction—and invoke the appropriate reasoning strategy. Because these rules are abstract, they are domain-independent; the logic used to resolve a conflict between two personal errands is the same logic applied to complex diplomatic negotiations.
Taxonomy of Goal Interactions
A central contribution of the paper is the categorization of how goals interfere with or complement one another. Wilensky identifies several key meta-situations:
- Goal Overlap: A state where a single action can satisfy multiple goals simultaneously, allowing for increased efficiency.
- Plan Non-integrability: A conflict where the execution of a plan for one goal inadvertently undoes a precondition for another goal.
- Recursive Subgoals: A failure state where a plan generates a subgoal identical to the original goal, leading to an infinite loop.
By making these interactions explicit, the meta-planning framework allows a system to prioritize and schedule tasks dynamically rather than following a static script.
Implementation and Results
The paper details the application of meta-planning in two distinct systems:
PAM (Plan Applier Mechanism)
PAM is a story-understanding program. By utilizing meta-planning, PAM can infer the motivations behind a character's "strange" or "efficient" behavior. If a character in a story performs an action that seems to contradict their primary goal, PAM uses meta-reasoning to determine if the character is resolving a conflict or exploiting a goal overlap, thereby producing a more sophisticated interpretation of the narrative.
PANDORA (Plan ANalyzer with Dynamic Organization, Revision and Application)
PANDORA is an autonomous planning program designed for situational reasoning. Unlike traditional planners that operate on a fixed set of instructions, PANDORA can detect new goals in real-time—such as a friend calling to visit—and use meta-planning to integrate these new objectives into its existing schedule. For example, it might decide to invite the friend over for a dinner it was already planning to cook, successfully identifying a goal overlap.
These implementations demonstrate that planning and understanding are symmetrical processes. A system capable of generating an intelligent plan for itself can use that same logic to explain the behavior of others.
Discussion and Limitations
While meta-planning represented a significant leap toward human-like reasoning, the paper acknowledges several limitations. As a conceptual and architectural proposal, it lacks a formal computational complexity analysis. There is little discussion on how these meta-rules scale as the number of goals increases, or how the system should choose between competing meta-plans when a conflict is truly unresolvable.
Despite these limitations, the work established a foundation for future research in multi-agent systems and dynamic replanning. It moved the field away from rigid, domain-specific scripts and toward a more flexible, declarative model of intelligence where the process of thinking is itself a subject of thought.
Summary updated