Centralized and Decentralized Federated Learning in Autonomous Swarm Robots: Approaches, Algorithms, Optimization Criteria and Challenges : The Sixth Edition of International Conference on Pattern Analysis and Intelligent Systems (PAIS’24)

Abstract

Fault tolerance and energy consumption optimization are critical issues in swarm robotics. This study examines recent approaches to address these challenges, focusing on a comparative analysis between Centralized Federated Learning (CFL) and Decentralized Federated Learning (DFL). CFL requires a centralized access point for model aggregation, while DFL eliminates the need for a central server, enabling aggregation at each node according to a specific architecture. The analysis of the results reveals that other approaches, such as Hybrid Federated Learning (HFL), more effectively meet the needs of intelligent agents (swarm robots). This effectiveness is particularly enhanced when HFL is combined with Deep Reinforcement Learning (DRL), resulting in Deep Hybrid Federated Reinforcement Learning (DHFRL). The results demonstrate that, although DFL eliminates the necessity of a central server, hybrid approaches are more efficient, especially when combined with Deep Reinforcement Learning (DRL), thus forming Deep Hybrid Federated Reinforcement Learning (DHFRL).