Federated Ensemble Model-Based Reinforcement Learning in Edge Computing

Federated learning (FL) is a privacy-preserving distributed machine paradigm that enables collaborative training among geographically and heterogeneous devices without gathering their data. Extending FL beyond the supervised models, federated reinforcement (FRL) was proposed to handle sequential decision-making problems in edge computing systems. However, existing FRL algorithms directly combine model-free RL with FL, thus often leading high sample complexity lacking theoretical guarantees. To address challenges, we propose novel algorithm effectively incorporates model-based ensemble knowledge distillation into for first time. Specifically, utilise create an of dynamics models clients, then train policy by solely using model interacting environment. Furthermore, theoretically prove monotonic improvement guaranteed. The extensive experimental results demonstrate our obtains much higher efficiency compared classic challenging continuous control benchmark environments under settings. also highlight significant impact client data local update steps on performance FRL, validating insights obtained from analysis.