2 August 2023 in IEEE Transactions on Wireless Communications
Network slicing at the radio access network (RAN) domain, called RAN slicing, requires elasticity, efficient resource sharing, and customization. In this scenario, radio resource scheduling (RRS) is responsible for dealing with scarce and limited frequency spectrum resources available at the RAN domain while fulfilling the slice intents. The wide variety of scenarios supported in 5G and beyond 5G networks makes the RRS problem in RAN slicing scenario a significant challenge. This paper proposes an intent-aware reinforcement learning method to perform the RRS function in a RAN slicing scenario. The slice’s quality of service intents is described in a common intent model in a service-level agreement. The proposed method tries to prevent intent faults by making the management of radio resources available among slices. This method uses slices’ and users’ equipment network metrics in the observation space. The proposed method is evaluated under different network conditions and outperforms different baselines considering the slices’ intents fulfillment.
24 November 2022 in IEEE Access
Fifth-generation (5G) cellular communication systems have embraced massive multiple-input-multiple-output (MIMO) in the low- and mid-band frequencies. In a multiband system, the base station can serve different users in each band, while the user equipment can operate only in a single band simultaneously. This paper considers a massive MIMO system where channels are dynamically allocated in different frequency bands. We treat multiband massive MIMO as a scheduling and resource allocation problem and propose deep reinforcement learning (DRL) agents to perform user scheduling. The DRL agents use buffer and channel information to compose their observation space, and the agent’s reward function maximizes the transmitted throughput and minimizes the packet loss rate. We compare the proposed DRL algorithms with traditional baselines, such as maximum throughput and proportional fairness. The results show that the DRL models outperformed baselines obtaining a 20% higher network sum rate and an 84% smaller packet loss rate. Moreover, we compare different DRL algorithms focusing on training time to assess the online implementation of the DRL agents, showing that the best agent needs about 50K training steps to converge.