Cultural Support for Improvisation

Philip E. Agre, Ian HorswillAAAI 1992

Abstract

In everyday environments, human agents perform complex sequences of actions with apparent ease. Traditional AI approaches often model these tasks as complex planning problems requiring extensive search. This paper argues that the computational burden of planning is mitigated by "cultural support"—the way our tools and environments are designed to simplify tasks. We present Toast, a system that leverages object-centered representations to perform kitchen tasks through improvisation, achieving performance orders of magnitude faster than classical planners.

Introduction: The Planning Dilemma

In early artificial intelligence research, a divide existed between Classical Planning and Reactive AI. Classical planners, such as STRIPS or SNLP, build exhaustive logical proofs of every step before execution. While mathematically rigorous, they are computationally expensive and often brittle. Conversely, reactive systems respond to immediate stimuli but frequently lack formal guarantees of goal achievement.

Agre and Horswill propose a middle ground: agents can act in real-time without a pre-computed plan if the environment is structured to support "improvisation." They argue that intelligence is not solely "in the head" but is distributed across cultural artifacts—tools evolved over centuries to make specific tasks backtrack-free.

The Computational Role of Culture

The core insight of this work is that human culture "pre-computes" solutions to common problems by embedding them into the design of tools. The authors categorize objects into two types:

  1. Tools: Objects that can always be returned to a "normal" state (e.g., a knife can be cleaned). These are used for "focused actions."
  2. Materials: Objects with acyclic state graphs (e.g., an egg that moves from fresh to beaten but never back).

By structuring the environment this way, most everyday tasks become "cooking tasks"—a formal class of problems where an agent can simply pick an unsatisfied goal and move a material one step closer to it using a "greedy" algorithm. If a tool is not ready, the agent performs a focused action to reset it. This eliminates the need for deep search or backtracking.

The Toast System

The authors implemented these principles in Toast, a system operating in a simulated kitchen with 110 objects. Instead of searching for a path, Toast uses "action tables"—essentially lookup sheets that dictate the next step based on the current state of a material and the desired goal.

A trace of Toast's execution showing interleaved tasks and goal completion.

As seen in the execution trace above, the system does not follow a rigid, linear plan. Instead, it opportunistically switches between tasks—such as starting toast while waiting for butter to melt, or setting the table while the toast is in the machine. This "improvisational" behavior emerges naturally from the greedy pursuit of multiple goals.

Results and Comparison

The efficiency of this approach is stark when compared to classical methods. Toast completed a complex breakfast task (omelette, pancakes, toast, and cleanup) in less than 30 seconds on a Macintosh IIx.

In contrast, when the authors attempted to solve the same problem using SNLP (a standard non-linear planner), the results were poor. While SNLP could handle simple toast, it failed to solve the omelette subproblem even after 6 hours of computation, eventually exhausting its memory. This demonstrates that for everyday tasks, the "heavy" logic of classical planning is often unnecessary and counterproductive.

Limitations and Future Work

While Toast excels at task sequencing, it faces challenges with "tending"—tasks that are time-sensitive, such as ensuring eggs do not burn while performing other actions. Additionally, the system can encounter resource deadlocks if it runs out of containers (like bowls), as containers are "committed" to a task until it is finished. Future work involves developing formal notions of priority and monitoring to handle these temporal constraints and physical limitations.

Conclusion

By externalizing complex logic into the design of our tools, culture allows us to internalize simpler, more efficient ways of acting. The Toast system demonstrates that when an environment is designed to support improvisation, agents can achieve complex goals with minimal computational overhead, suggesting a new path for developing robust, real-world AI.

Summary updated

BibTeX
@inproceedings{agre1992cultural,
  author = {Philip E. Agre and Ian Horswill},
  title = {{Cultural Support for Improvisation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1992},
}
Figure