Deep Reinforcement Learning for Power control in Multi-tasks Wireless Cellular Networks

Extensive spectrum reuse due to the massive deployment of small base stations will incur high interference in future networks. One solution to deal with this interference is implementing power control schemes. However, solving the power allocation in denser wireless networks becomes challenging due to the computational complexity involved and the complete Channel State Information (CSI) required. One alternative to address these challenges is implementing Deep Reinforcement Learning (DRL) models, which may find solutions faster and with fewer CSI requirements. In this work, we implement a Centralized Training scheme for training a DRL model over a wireless network environment consisting of different tasks (i.e., UE location and CSI). Training over various tasks may degrade the performance of older tasks. Therefore, to retain the knowledge of older tasks, we implement a memory-based strategy to split the buffer for storing relevant experiences based on the action-space exploration. Additionally, we implement a rehearsal method for reusing samples of similar environments. The results show a learning performance improvement when learning newer tasks. Further, the degradation due to learning multi-task is reduced compared to classic DQN models with FIFO buffer.