On Waiting

Arthur M. FarleyAAAI 1980

Abstract

In the early development of artificial intelligence, problem-solving systems typically operated under "closed-world" assumptions where the agent was the sole actor in a static environment. Arthur M. Farley’s 1980 paper, "On Waiting," challenges this paradigm by formalizing waiting as a strategic choice rather than a passive delay. By defining waiting as the active maintenance of preconditions for future actions, Farley provides a framework for AI agents to navigate real-world constraints such as schedules, external collaborators, and environmental factors.

Introduction

Traditional AI planners often viewed unmet preconditions as failures requiring backtracking or immediate corrective action. However, in dynamic environments populated by other agents and governed by temporal schedules, the most efficient strategy is often to wait. Whether waiting for a store to open or for a collaborator to provide a tool, an agent must be able to reason about when to wait and how to utilize that time effectively. Farley’s work seeks to categorize these scenarios and integrate them into a formal problem-solving model.

Methodology: Partitioning the World

Farley extends knowledge-based problem-solving models by partitioning the propositions (facts) of a plan state into three distinct components based on who or what controls them:

  1. SELFOP: Conditions the system can satisfy through its own operators.
  2. OTHEROP: Conditions expected to be satisfied by other agents or external systems.
  3. NOOP: Conditions outside the control of any agent, such as the passage of time or weather events.

By evaluating the satisfaction status of these components, the system can classify its current state and determine if waiting is the appropriate course of action.

Taxonomy of Waiting

The core of Farley's framework is a taxonomy of rule classifications that trigger the decision to wait. This logic allows an AI to distinguish between being ready to act and being constrained by external factors.

Table I: Waiting Rule Classification based on component satisfaction.

As shown in Table I, several key classes emerge:

  • Ready: All components (SELFOP, OTHEROP, NOOP) are satisfied (S), allowing for immediate action.
  • Imminent: The agent and other actors are ready, but a temporal constraint (NOOP) is expected (E) but not yet met.
  • Dependent: The agent is ready, but is waiting for an external agent (OTHEROP) to satisfy a condition (N).
  • Needy: The agent is waiting for both an external agent and a specific time to align.

Strategic Waiting and Parallelism

A significant insight of the paper is that waiting is not synonymous with inactivity. Farley introduces the concept of active waiting, where an agent can execute compatible rules—tasks that can be completed during the waiting period without disrupting the preconditions being maintained for the primary goal.

Conversely, the agent must identify contending rules. These are actions that, if taken, would destroy the conditions currently being maintained, potentially leading to plan failure. This formalization allows for more sophisticated multi-goal planning where trade-offs between immediate progress and long-term maintenance are explicitly weighed.

Conclusion and Future Directions

While the 1980 paper lacks empirical data or modern benchmark testing, it established a foundational logic for reasoning about time and external agency in AI. Farley acknowledged that the model depends heavily on accurate heuristic estimates of time and the reliability of external actors. Future work was proposed to involve simulations of complex urban environments to evaluate different waiting policies. Despite its age, the paper’s definition of waiting as "maintenance of state" remains a relevant perspective for modern autonomous systems operating in collaborative and time-sensitive domains.

Summary updated

BibTeX
@inproceedings{farley1980waiting,
  author = {Arthur M. Farley},
  title = {{On Waiting}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1980},
}

Loading related content...

Figure