A Logic of Situated Know-How

Munindar P. SinghAAAI 1991

Introduction

In the early 1990s, the prevailing paradigm in artificial intelligence research was heavily centered on "know-that"—the notion that intelligence is primarily a knowledge base of facts coupled with a logical interpreter. Under this view, an agent’s ability to act was often reduced to a subset of propositional knowledge: an agent "knows how" to perform an action if it possesses a plan and "knows that" the plan will succeed.

In "A Logic of Situated Know-How," Munindar Singh challenges this reductive view. He argues that knowing the physics of a tennis swing (factual knowledge) is fundamentally different from the ability to play tennis (skill). This paper introduces a formal mathematical framework to treat "know-how" as a distinct logical category, providing sound and complete axiomatizations for agents operating in complex, unpredictable environments.

The Problem of Situatedness

Traditional theories of action often assumed that agents have the luxury of complete information and infinite time to formulate explicit plans. However, real-world agents are "situated." They must act reactively, often with incomplete data and under strict temporal constraints.

Singh identifies a critical gap in AI theory: the lack of a formal logic that describes these skills without forcing them into the mold of factual knowledge. A situated agent might not have a complete, explicit plan before it begins to move; rather, its know-how is expressed through its ability to respond to the environment as it evolves. To capture this, the paper moves away from static plans and toward a dynamic model of agent behavior.

Formal Methodology

The paper constructs its framework by synthesizing Branching-Time Logic (CTL*) and Dynamic Logic (DL). This combination allows for the modeling of both the passage of time and the consequences of specific actions within a single system.

Scenarios and Action Trees

The world is modeled as a branching history of possible futures. Within this structure, "know-how" is defined relative to a "tree of actions." This tree does not necessarily represent a conscious plan held in the agent's memory; instead, it represents the agent’s selection function—the mechanism by which it chooses the next action based on the current state of the environment.

Two Levels of Know-How

Singh bifurcates the concept of skill into two distinct layers:

  1. Reactive Know-How ($K'$): A bottom-up definition where an agent knows how to achieve a goal if it can "force" the environment into a state where the goal is met, regardless of how the environment branches or reacts.
  2. Strategic Know-How: A higher-level abstraction where behavior is described using complex strategies. These strategies function similarly to computer programs, employing operators such as if-then, while loops, and sequences to compose simpler reactive skills into complex behaviors.

Key Insights and Logical Results

Non-Reductivity

A central thesis of the work is that know-how is a fundamental category. The essence of a skill is lost when one attempts to translate it into a list of "know-that" facts. Skill is defined by the interaction between the agent and a changing environment, not just the internal state of the agent's database.

The Paradox of Achievement

Singh introduces a subtle distinction between simple ability ($K'$) and the ability to bring something about ($K$). This leads to a unique logical result: if a goal $p$ is already inevitable or currently true, an agent technically does not "know how to achieve" it in the strategic sense. To "know how" implies the power to change the world to meet a condition that might not otherwise be met.

Soundness and Completeness

The core contribution of the paper is the development of logical axioms for these definitions. Singh provides mathematical proofs for two major theorems:

  • Theorem 1: Establishes a sound and complete axiomatization for reactive know-how, providing a toolkit for proving what an agent can or cannot achieve within a given model.
  • Theorem 2: Extends this to strategic know-how, proving that the logic remains robust even when describing complex, program-like agent behaviors.

Implications for AI Design

The strategic model suggests a two-layered architecture for intelligent systems. In this design, a top layer handles the composition of complex tasks (strategic), while a bottom layer manages the reactive, "muscle memory" execution of those tasks. This allows designers to reason about "macro-operators," simplifying the analysis of complex systems by grouping atomic actions into meaningful, provable strategies.

Limitations and Future Directions

As a foundational theoretical work, the paper focuses primarily on single-agent environments and discrete actions. The author acknowledges that the continuous nature of physical skills and the complexities of multi-agent coordination remain areas for further exploration. However, by providing a rigorous logic for situated know-how, Singh provides the necessary groundwork for formalizing how agents communicate intentions and coordinate actions in the real world.

Summary updated

BibTeX
@inproceedings{singh1991logic,
  author = {Munindar P. Singh},
  title = {{A Logic of Situated Know-How}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1991},
}

Loading related content...

Figure