Efficient Load Balancing for Future Dense Networks using Radio over Fiber Infrastructure and applying Different Learning Rates

Reinforcement Learning (RL) can lead to effective Load-Balancing (LB) mechanisms, as traditional methods cannot always provide an optimal solution in cellular networks. This study proposes an RL-based LB scheme for a dense network that uses radio over fiber infrastructure. The proposed technique is based on LB constraints in the action space that maintain zero violation during the learning process. In this technique, a Deep Q-Network agent was chosen to search for an optimal policy to maximize the expected cumulative long-term reward to satisfy the constraints. This study uses the number of user entities per base station in the dense network as constraints to maintain average throughput based on the Signal-to-Noise Ratio (SNR) generated from the radio frequency signals of the network. The proposed method outperformed at an SNR of 38 dB with a throughput of 32 Mbps for a 20 MHz channel bandwidth for macro- and microcells in the dense network. Furthermore, this study examined the effect of different learning rates as hyperparameters in the system. The proposed approach shows that when the agent was trained with a learning rate of 1e-3, the network performed well by obtaining a higher CDF compared to a learning rate of 1e-5. In addition, the system achieved higher rewards for a learning rate of 1e-3 with or without LB constraints, confirming the efficiency of the proposed scheme. The simulation results showed that CDF was 4% higher when using constraints compared to without constraints.