Wireless networks are and will continue to be a fundamental cornerstone of the global digital economy and our connected society. Meanwhile, the fast traffic growth and demand diversification challenges are driving wireless networks to evolve. To cope with the estimated 1000x traffic growth, the fifth generation (5G) and beyond 5G will be hyper-dense heterogeneous networks (HetNet), featuring the coexistence of different radio access technologies (RATs) such as LTE/LTE-Advanced, WiFi, Internet of Things (IoT), and 5G New Radio (NR) on terrestrial and aerial platforms. Diverse data service demand has led to network function virtualization (NFV) implementation. NFV slices (connected cars, mMTC, URLLC, IoT, etc.,) will have resource parameters optimized according to different service requirements. Due to the aforementioned changes, it is envisaged that 5G & Beyond radio resource management (RRM) will be significantly more complex, with a dimensionality explosion in adjustable parameters.
To tackle this challenge, deep reinforcement learning (DRL) has been recently introduced in RRM. Thanks to its powerful representation capability, deep neural networks (DNN) is able to approximate policy or value functions for large-scale Markov decision process (MDP) models, which are intractable for traditional tabular-based RL algorithms. Whilst the issues of model over-fitting, robustness to data quality, and brittleness to malicious data remain open issues, we expect AI to transform a number of solutions to challenging wireless communication optimisation problems. Indeed, the explosive growth in AI for wireless communication research indicates this. However, the adoption of DNN risks increasing the energy consumption of the network significantly, which is receiving increasing attention in the whole AI community. It is reported that the computations required for deep learning (DL) research have an estimated 300,000x increase from 2012 to 2018, which results a surprisingly large carbon footprint . Recent study revealed that the estimated CO2 emissions from training even could be larger than that of a car in a lifetime including fuel consumption. Realizing such energy consumption crisis, researchers are now appealing that we should prioritize computation and energy efficiency and design more environmentally friendly and inclusive DL algorithms. Considering the long-term running and vast deployment nature of resource and network management, if the performance improvement comes at the price of excessive energy consumption that even erodes the gain, DRL can be unacceptable as a sustainable solution. Therefore, realizing green DRL wireless communication algorithms is a global priority.
Specific Objectives: To achieve energy-efficient DRL for customized RRM applications, the 2-year fellowship will address current limitations via the following objectives (O1 & O2):
1. O1- Efficient Distributed DRL Architecture: Enhance the scalability and efficiency of DRL running architecture for decoupling the computation-intense and large-scale RRM task through a distributed manner.
2. O2- Energy-Efficient DRL Algorithms: Compress the whole model of DRL to jointly reduce the energy consumption of the training and inference stages.
