Model-based reinforcement learning approach for federated learning resource allocation and parameter optimization

In this paper, we investigate the performance of a model-based approach for solving resource allocation and parameter adjustment problems in federated learning (FL) within a wireless network. Given the existence of models for energy, communication channels, and accuracy, such models can be leveraged to achieve improved performance. Additionally, machine learning techniques can be employed to identify known parts of the model and also exploit training data for unknown parts of the model, enabling the creation of complex policies. Model-based reinforcement learning (RL) methods have the potential to offer such solutions, particularly in resource allocation and parameter optimization settings where the model can be partially derived mathematically. Our results demonstrate that the use of such a method in FL scenarios leads to improvements in both performance and the number of iterations required to identify the desired policy. Our simulations demonstrate the significance of allocating appropriate resources for FL applications through proper consideration of inherent tradeoffs, as performance will not improve beyond a certain saturation point. Additionally, our proposed FL model takes intelligently into account the presence of slow users to propose efficient policies for users that may have access to more abundant resources.