Self-Optimized Agent for Load Balancing and Energy Efficiency: A Reinforcement Learning Framework With Hybrid Action Space

We consider the problem of jointly enhancing the network throughput, minimizing energy consumption, and improving the network coverage of mobile networks. The problem is cast as a reinforcement learning (RL) problem. The reward function accounts for the joint optimization of throughput, energy consumption, and coverage (through the number of uncovered users); our RL framework allows the network operator to assign weights to each of these cost functions based on the operator's preferences. Moreover, the state is defined by key performance indicators (KPIs) that are readily available on the network operator side. Finally, the action space for the RL agent comprises a hybrid action space, where we have two continuous action elements, namely, cell individual offsets (CIOs) and transmission powers, and one discrete action element, which is switching MIMO ON and OFF. To that end, we propose a new layered RL agent structure to account for the agent hybrid space. We test our proposed RL agent over two scenarios: a simple (proof of concept) scenario and a realistic network scenario. Our results show significant performance gains of the proposed RL agent compared to baseline approaches, such as systems without optimization or RL agents that optimize only one or two parameters.'