Learning to Solve Decision Problems over Two Timescales: an Application to 5G Puncturing.

One of the biggest innovations on 5G and beyond is the support of three different services with particular delay and bandwidth requirements, such as Massive Machine Type Communications (MMTC), enhanced Mobile Broad Band (eMBB) and Ultra-Reliable Low Latency Communications (URLLC). In order to achieve these multiple service requirements, all users have to share resources over the 5G Orthogonal Frequency-Division Multiple Access (OFDMA) frame. One of the strategies proposed by the 5G standard is puncturing, which allows the scheduler to assign eMBB services on a timescale, and on a shorter timescale to preemptively overwrite part of the eMBB assignment when a URLLC user arrives. The optimization of puncturing poses a challenging problem: the optimal allocation depends on traffic arriving over different timescales, which forces the scheduler to make allocation decisions without knowledge of future users’ demands, all while having to satisfy several strong constraints. This kind of multiple timescales optimization with restrictions is also to be found in many interesting problems, such as energy management. We propose a learning mechanism where the system learns offline the optimal allocation according to the network state. This learned estimation is then used online to determine the optimal allocation. Through simulations, we verify that the proposed learning strategy provides results close to the optimal policy, improving state of the art proposals for puncturing schemes.