Cost-efficient Federated Reinforcement Learning- Based Network Routing for Wireless Networks

Advances in Artificial Intelligence (AI) provide new capabilities to handle network routing problems. However, the lack of up-to-date training data, slow convergence, and low robustness due to the dynamic change of the network topology, makes these AI-based routing systems inefficient. To address this problem, Reinforcement Learning (RL) has been introduced to design more flexible and robust network routing protocols. However, the amount of data (
$i$
. e., state-action space) shared be- tween agents, in a Multi-Agent Reinforcement Learning (MARL) setup, can consume network bandwidth and may slow down the process of training. Moreover, the curse of dimensionality of RL encompasses the exponential growth of the discrete state-action space, thus limiting its potential benefit. In this paper, we present a novel approach combining Federated Learning (FL) with Deep Reinforcement Learning (D RL) in order to ensure an effective network routing in wireless environment. First, we formalize the problem of network routing as a problem of RL, where multiple agents that are geographically distributed train the policy model in a fully distributed manner. Thus, each agent can quickly obtain the optimal policy that maximizes the cumulative expected reward, while preserving the privacy of each agent's data. Experiments results show that our proposed Federated Reinforcement Learning (FRL) approach is robust and effective.