Analysis of Evolved Response Thresholds for Decentralized Dynamic Task Allocation

TL;DR

This paper uses a multi-objective genetic algorithm to evolve response thresholds in a decentralized swarm, significantly improving task allocation performance and generalizing across different problem instances.

Contribution

It introduces a method to evolve universal response thresholds for swarm task allocation, outperforming traditional and dynamic thresholds, and applicable across multiple problem scenarios.

Findings

Evolved thresholds outperform uniform and dynamic thresholds.

Thresholds generalize across different task instances.

Evolved thresholds nearly achieve optimal performance.

Abstract

We investigate the application of a multi-objective genetic algorithm to the problem of task allocation in a self-organizing, decentralized, threshold-based swarm. Each agent in our system is capable of performing four tasks with a response threshold for each, and we seek to assign response threshold values to all of the agents a swarm such that the collective behavior of the swarm is optimized. Random assignment of threshold values according to a uniform distribution is known to be effective; however, this method does not consider features of particular problem instances. Dynamic response thresholds have some flexibility to address problem specific features through real-time adaptivity, often improving swarm performance. In this work, we use a multi-objective genetic algorithm to evolve response thresholds for a simulated swarm engaged in a dynamic task allocation problem: two-dimensional collective tracking. We show that evolved thresholds not only outperform uniformly distributed thresholds and dynamic thresholds but achieve nearly optimal performance on a variety of tracking problem instances (target paths). More importantly, we demonstrate that thresholds evolved for one of several problem instances generalize to all other problem instances eliminating the need to evolve new thresholds for each problem to be solved. We analyze the properties that allow these paths to serve as universal training instances and show that they are quite natural.