Risk-Aware Task Migration for Multiplex Unmanned Swarm Networks in Adversarial Environments

Kai Di, Tienyu Zuo, Pan Li, Yuanshuang Jiang, Fulin Chen, Yichuan JiangIJCAI 2025 · DOI

TL;DR

This paper introduces a novel "Two-Scale Adversarial Task Migration" (TSATM) algorithm designed to dynamically reassign tasks among unmanned swarm agents operating in complex, adversarial environments. By considering both individual agent risks and overall network layer vulnerabilities, TSATM significantly improves task completion rates and system robustness compared to existing methods, while maintaining computational efficiency.

What Problem Does This Paper Address?

Autonomous unmanned swarms are increasingly deployed in challenging, adversarial missions where they must simultaneously perform diverse tasks like reconnaissance, communication, and combat. To achieve this, swarms naturally form "multiplex network structures," where agents dynamically participate in multiple coordination layers. However, the dynamic nature of these environments, coupled with frequent task variations, leads to imbalanced workloads across agents and network layers. This imbalance increases the risk of enemy detection and destruction, a critical issue that current task scheduling approaches, which often simplify these complex networks into single layers, fail to adequately address. Designing effective task migration strategies for these multiplex networks under dynamic, adversarial conditions is an NP-hard problem, necessitating a more sophisticated approach.

Methods

The core contribution of this paper is the Two-Scale Adversarial Task Migration (TSATM) algorithm. It employs a hierarchical, two-scale approach:

  • Inter-layer Task Migration: This phase focuses on global load balancing. It identifies "key agents" across different network layers based on operational reliability, centrality, and multi-layer connectivity. These key agents then facilitate information exchange and task redistribution between layers. A sophisticated fitness comparison function is used to determine the suitability of a target layer for task immigration, considering migration cost, task requirement satisfaction, and current layer workload.
  • Intra-layer Task Migration: This phase handles local optimization within a specific network layer. It considers agent load, task execution capability, and migration cost to determine the best agent for task acceptance. An agent task acceptance capability comparison function guides this process, ensuring that tasks are migrated to agents that can best handle them without significantly increasing their own or the layer's risk.

TSATM combines these two scales to achieve both global equilibrium and local efficiency, addressing the limitations of single-layer approaches.

Datasets

Not discussed in the paper. The experiments were conducted using a simulated framework with varying parameters for network layers, agents, tasks, and dynamic change rates.

Key Insights and Ideas

  • Multiplex Networks are Key: The paper highlights that understanding and leveraging the multiplex nature of swarm networks is crucial for effective task management in adversarial settings.
  • Risk is Multi-faceted: Simply balancing load isn't enough. The approach correctly identifies and quantifies both individual agent-level risk (detection due to high load) and layer-level risk (implication and destruction of the entire layer).
  • Two-Scale Approach is Superior: The combination of inter-layer (global) and intra-layer (local) migration strategies is shown to be significantly more effective than single-scale approaches, achieving better load balancing and risk mitigation.
  • Key Agents as Information Hubs: The concept of "key agents" acting as central points for inter-layer communication and coordination is a practical and effective way to manage complexity.
  • Heuristic Efficiency: TSATM demonstrates that a well-designed heuristic approach can achieve near-optimal results with significantly lower computational overhead than exact solvers like CPLEX, making it practical for real-time deployments.

Key Results

The TSATM algorithm consistently outperformed benchmark algorithms (COTM, CARC) across various experimental scenarios:

  • Task Numbers: TSATM showed superior performance in objective function value and task completion cost, especially under high task loads. It maintained higher expected task completion proportions, closely approximating the optimal CPLEX solution.
  • Agent Numbers: TSATM demonstrated robust scalability, with its objective function value showing a positive correlation with agent numbers before plateauing. It consistently outperformed benchmarks and approached the CPLEX solution.
  • Task Load Change Rates: TSATM exhibited superior stability and optimal performance under increasing dynamic task load variation rates. It showed minimal cost escalation and significantly less degradation in task completion proportions compared to other algorithms.
  • Computational Efficiency: TSATM maintained execution times below 80ms, even under highly dynamic conditions, making it practical for real-world deployments, unlike the exponential complexity of CPLEX.

Limitations and Future Work

  • Limitations: While the paper demonstrates strong performance, the complexity of real-world adversarial environments might present further challenges not fully captured in the simulation. The paper also acknowledges that the optimal solution is computed by CPLEX, which is computationally expensive, implying that TSATM, while efficient, might not always reach the absolute theoretical optimum.
  • Future Work: The authors suggest that future work could explore more sophisticated agent selection mechanisms and investigate the impact of different network topologies and communication constraints on the TSATM algorithm. Further research into adaptive weighting of objective function parameters based on real-time environmental feedback could also enhance performance.

Summary updated

BibTeX
@inproceedings{di2025risk,
  author = {Kai Di and Tienyu Zuo and Pan Li and Yuanshuang Jiang and Fulin Chen and Yichuan Jiang},
  title = {{Risk-Aware Task Migration for Multiplex Unmanned Swarm Networks in Adversarial Environments}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/971},
}
Figure