[論文レビュー] Building Large-Scale Drone Defenses from Small-Team Strategies

Grant Douglas, Stephen Franklin+2 more|arXiv (Cornell University)|Feb 13, 2026

UAV Applications and Optimization被引用数 0

ひとこと要約

The paper presents a hybrid GA–DP framework that builds large-scale drone defenses by reusing small-team heuristics, assembled via dynamic programming, with iterative refinement and LLM-assisted heuristic generation. It demonstrates scalability to attacker swarms up to 30 and defender swarms up to 45.

ABSTRACT

Defending against large adversarial drone swarms requires coordination methods that scale effectively beyond conventional multi-agent optimisation. In this paper, we propose to scale strategies proven effective in small defender teams by integrating them as modular components of larger forces using our proposed framework. A dynamic programming (DP) decomposition assembles these components into large teams in polynomial time, enabling efficient construction of scalable defenses without exhaustive evaluation. Because a unit that is strong in isolation may not remain strong when combined, we sample across multiple small-team candidates. Our framework iterates between evaluating large-team outcomes and refining the pool of modular components, allowing convergence on increasingly effective strategies. Experiments demonstrate that this partitioning approach scales to substantially larger scenarios while preserving effectiveness and revealing cooperative behaviours that direct optimisation cannot reliably discover.

研究の動機と目的

• Motivate scalable defense against adversarial drone swarms beyond small-team optimization.
• Develop a modular framework that reuses small-team heuristics as building blocks for larger forces.
• Enable tractable scaling through dynamic programming and chromosome factorisation.
• Incorporate heuristic generation, including LLM-assisted candidates, to expand the search space.
• Demonstrate iterative refinement to converge on high-performing large-scale defense strategies.

提案手法

• Hybrid GA–DP framework that evolves small-team chromosomes representing heuristic, spawn location, and parameters.
• Dynamic programming to allocate evolved small-team components into larger defender configurations.
• Stage-wise pipeline: GA evolution (Stage 1), DP-based allocation (Stage 2), chromosome sampling (Stage 3), iterative refinement (Stage 4).
• Chromosome-based hierarchical policy where each agent’s low-level behavior is defined by a chosen heuristic.
• Use of LLM-generated heuristics to diversify candidate strategies.
• Implementation in JAX with high-performance computing to simulate billions of time steps.

実験結果

リサーチクエスチョン

• RQ1 Can small-team evolved heuristics be effectively reused to defend much larger swarms?
• RQ2 Does integrating DP with GA and heuristic factorisation yield scalable, high-performing defense strategies?
• RQ3 What is the impact of iterative refinement on large-scale defense performance?
• RQ4 Are chromosome-level co-dependencies important for effective large-scale strategies?
• RQ5 How does LLM-assisted heuristic generation influence search efficiency and outcomes?

主な発見

• Stage 1 GA evolution improves small-scale defender performance compared to random initialization.
• Stage 2 DP-informed chromosome allocation enables scalable composition for larger engagements.
• Stage 3 chromosome sampling with priors substantially outperforms GA baseline in large-scale scenarios.
• Stage 4 iterative refinement yields the highest and most consistent win rates across swarm sizes.
• Full pipeline reveals cooperative behaviors and synergistic heuristics that single-stage approaches miss.
• The approach scales to attacker swarms up to 30 and defender swarms up to 45 with stable execution.