Decentralized adaptive task allocation for dynamic multi-agent systems.

Area of Science:

• Artificial Intelligence
• Robotics
• Distributed Systems

Background:

• Multi-agent systems require efficient dynamic task assignment strategies.
• Existing methods often struggle with partial observability, noisy feedback, and limited communication.

Purpose of the Study:

• To develop a decentralized two-layer architecture for dynamic task assignment.
• To enable scalable, online task allocation without centralized coordination.
• To ensure robustness under challenging operational conditions.

Main Methods:

• Adaptive controllers using recursive regression with forgetting for parameter prediction.
• Selective task broadcasting based on relevance and agent availability.
• Distributed optimization combining Simultaneous Perturbation Stochastic Approximation (SPSA) and consensus-based synchronization.

Main Results:

• Demonstrated scalability and online task allocation capabilities.
• Showcased robustness across varying noise levels, task dynamics, and input arrival patterns.
• Successfully applied to prompt-based task assignment for large language models (LLMs).

Conclusions:

• The proposed decentralized architecture effectively addresses dynamic task assignment in multi-agent systems.
• The system provides a robust and scalable solution for real-world applications.
• This approach facilitates efficient coordination in complex, partially observable environments.