Fitness-Based Dynamic Leader Election for Heterogeneous Autonomous Agent Swarms: An Information-Theoretic Approach

Leader election is a foundational primitive in distributed systems, yet existing algorithms — including Bully, Ring, Raft, and Paxos — assume relatively homogeneous nodes and stable network conditions. These assumptions break down in heterogeneous autonomous agent swarms operating in dynamic, contested, or partially connected environments, where the optimal leader at any moment depends on a combination of computational capacity, sensor coverage, network connectivity, energy reserves, and task-specific fitness. This paper introduces a Fitness-Based Dynamic Leader Election algorithm grounded in an information-theoretic fitness function that quantifies each agent's suitability to lead at any given moment based on a multi-dimensional state vector. The algorithm enables continuous, decentralized re-election as conditions change, without requiring global synchronization or fixed leader terms. We evaluate the algorithm against five baseline approaches (Bully, Ring, randomized election, capacity-based, and proximity-based) across simulated swarms of 10 to 100 heterogeneous agents under varying failure rates, mobility patterns, and network topologies. Results demonstrate that fitness-based election achieves 30 to 70 percent improvement in mission continuity and decision quality compared to baselines, particularly in adversarial and partially connected scenarios. The algorithm is designed for deployment in autonomous drone swarms, distributed sensor networks, and edge AI systems where leadership must adapt rapidly to changing operational conditions. This work forms one component of a broader federated drone swarm intelligence architecture (USPTO Provisional Patent Application #64/002,060).