Experimentally Evaluating Communicative Strategies: The Effect of the Task

Marilyn A. WalkerAAAI 1994

Introduction

In the fields of artificial intelligence and linguistics, the "efficiency hypothesis" has long posited that the most effective communication is the most concise, minimizing the number of utterances. However, natural human dialogue often contains "Informationally Redundant Utterances" (IRUs)—statements that repeat known information or make obvious inferences explicit. Marilyn Walker’s research challenges the efficiency hypothesis by demonstrating that redundancy is a functional strategy for resource-bounded agents, helping them manage memory limitations and coordinate successfully in complex, fault-intolerant tasks.

The Design-World Environment

To empirically test the impact of communicative strategies, the study employs "Design-World," a collaborative task environment. In this simulation, two agents must work together to arrange furniture in a two-room house. The agents' objective is to achieve the highest possible utility score, which is calculated based on the point values assigned to different pieces of furniture.

The Design-World collaborative task environment, where agents arrange furniture to maximize utility.

Methods: Modeling Resource-Bounded Agents

The agents in this study are built using the IRMA architecture, augmented with a model of Limited Attention/Working Memory (AWM). This model treats memory as a chronological 3D space where the ability to retrieve beliefs is constrained by a "search radius" or Hamming distance.

The agent architecture, highlighting the central role of Attention/Working Memory (AWM) in reasoning and deliberation.

As shown in the architecture, the AWM interacts directly with belief deliberation and the means-ends reasoner. By varying the search radius, the researchers can simulate agents with different cognitive capacities—from severely limited memory (radius of 1) to near-perfect recall (radius of 16). This architectural constraint is what makes redundant strategies necessary; they effectively "pull" information back into the agent's current search radius, acting as a rehearsal mechanism.

Communicative Strategies and Task Constraints

The study compares three distinct strategies:

  • All-Implicit: The most concise approach, providing only necessary new information and avoiding any redundancy.
  • Close-Consequence: Agents explicitly state the effects of an action (e.g., "Since we used the green rug, you no longer have it") immediately after an agreement is reached.
  • Explicit-Warrant: Agents provide the "why" (the utility or point value) of a proposal immediately, even if that information was previously shared.

The effectiveness of these strategies is highly dependent on the task type. In "Standard Tasks," where errors merely reduce the score, redundancy can be detrimental by displacing other useful facts from the agent's limited memory. However, in "Zero-Invalid" tasks—where a single error, such as reusing a furniture item, leads to total mission failure—redundancy acts as a critical safeguard.

Key Findings: The Efficiency Paradox

The research reveals an "Efficiency Paradox": the most concise strategy (All-Implicit) often performs the worst in high-stakes or high-coordination scenarios.

  1. Salience Management: IRUs serve as a "rehearsal" mechanism, keeping critical facts within the agent's limited search radius so they are not forgotten during complex reasoning.
  2. Fault Tolerance: In fault-intolerant tasks, the Close-Consequence strategy significantly improves performance for agents with mid-range memory limits (AWM 3 to 6) by preventing the accidental reuse of resources.
  3. Cognitive Cost Reduction: The Explicit-Warrant strategy is essential when tasks require high levels of belief coordination. By providing the rationale for a proposal upfront, it saves agents the cognitive cost of searching their own memory for the utility of a piece of furniture.

Conclusion

Walker’s work suggests that communicative "efficiency" should not be measured solely by the number of utterances, but by the success of the collaborative goal given the agents' cognitive constraints. In complex, real-world-like scenarios, being "wordy" is often more effective than being concise. Informationally redundant utterances are not signs of human fallibility but are vital tools for robust coordination in resource-bounded systems.

Summary updated

BibTeX
@inproceedings{walker1994experimentally,
  author = {Marilyn A. Walker},
  title = {{Experimentally Evaluating Communicative Strategies: The Effect of the Task}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1994},
}
Figure