Using Joint Responsibility to Coordinate Collaborative Problem Solving in Dynamic Environments

Nicholas R. Jennings, E. H. MamdaniAAAI 1992

Abstract

This paper addresses the "complexity barrier" in multi-agent systems (MAS) by proposing "Joint Responsibility," a behavioral framework for coordinating autonomous agents in dynamic environments. Unlike previous abstract theories of joint intentions, Joint Responsibility provides a concrete code of conduct for agents when shared plans fail. Evaluated in a simulated electricity transportation management domain, the framework demonstrates a significant reduction in wasted computational effort compared to selfish or emergent coordination strategies, albeit with increased communication overhead.

Introduction

In the early 1990s, researchers recognized that building monolithic AI systems to solve complex industrial problems, such as national power grid management, was increasingly impractical. The alternative—Multi-Agent Systems (MAS)—offered modularity but introduced a coordination crisis. If one agent in a team fails or finds a task impossible, how should it signal this to others to prevent wasted resources? Existing theories defined what it meant for a group to have a goal but lacked a practical "instruction manual" for agent behavior during environmental shifts or plan failures.

Jennings and Mamdani (1992) tackle this by introducing "Joint Responsibility," a meta-level framework that provides a social contract for cooperating agents. By mandating communication when commitments are dropped, the framework aims to bridge the gap between individual autonomy and collective efficiency.

The Joint Responsibility Framework

Joint Responsibility moves beyond abstract mental states to a concrete "prescription of behavior." The framework defines strict preconditions for collaboration, including identifying participants and agreeing on a common objective and code of conduct.

Under this prescription, if an agent can no longer commit to a goal—whether because the goal is met, it has become impossible, or the motivation has vanished—it cannot simply quit. Instead, it must:

  1. Suspend local actions.
  2. Determine if a local fix is available.
  3. Inform all team members of the failure and the specific reason why.

This logic was implemented in the GRATE (General Reasoning Architecture for multi-agent-TEams) system. GRATE separates the "Domain Level System," which contains the problem-solving code, from the "Cooperation and Control Layer," the "social brain" that manages commitments and team dynamics.

Experimental Evaluation

The authors evaluated the framework using a simulated industrial domain: Electricity Transportation Management. The simulation involved three specialized agents: CSI (Detects faults), AAA (Pinpoints faulty elements), and BAI (Identifies out-of-service areas).

The multi-agent system for electricity transportation management, featuring CSI, AAA, and BAI agents.

The environment was designed to be unpredictable, with transient faults and changing network topologies. The performance of "Responsible" agents was compared against "Selfish" agents (who quit without notice) and "Implicit" (emergent) groups across 100 simulation runs.

Key Results

The primary metric for success was "Average Wasted Effort," defined as the time elapsed between a failure occurring and the agents stopping work on the obsolete task.

Average wasted effort versus the probability of a joint action becoming unsustainable.

As the environment became more unstable (a higher chance of plans becoming unsustainable), the Responsibility framework showed significantly lower wasted effort compared to the other models. Surprisingly, this increased logical complexity did not require significantly more processing power, though it did result in higher communication overhead due to the constant status updates between agents.

Discussion and Future Work

The core insight of this research is that "selfishness" in a collaborative system—quitting a task without informing others to save local CPU cycles—is globally expensive. By treating intentions as "conduct controllers," the authors provided an implementable logic for engineering robust multi-agent teams.

However, limitations remain. The "chattiness" of the protocol could become a bottleneck in extremely time-critical environments. Future work suggests using case-based reasoning to speed up coordination and exploring how this code of conduct scales to larger communities of hundreds or thousands of agents.

Summary updated

BibTeX
@inproceedings{jennings1992using,
  author = {Nicholas R. Jennings and E. H. Mamdani},
  title = {{Using Joint Responsibility to Coordinate Collaborative Problem Solving in Dynamic Environments}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1992},
}
Figure