The 5G cellular network standards specify the concept of network slicing
that allows diverse types of network services to run on the underlying
physical network infrastructure. Network Slicing in 5G networks enables
allocation and sharing of the underlying network resources of an
infrastructure provider (INP) among multiple tenants of the INP. A slice
is defined as an end-to-end independent virtual logical network assigned
resources from three 5G network domains, namely Radio Access Network
(RAN), Transport, and Core. Network resources from these three domains
are virtualized and shared among the other slices running while
maintaining isolation. Each tenant generates slice requests specifying
the resources required, and the INP collects revenue from the tenants
for hosting the admitted slices. The slices are classified as elastic
and inelastic, with flexible and rigid service guarantee requirements,
respectively.

The INP uses a Slice Admission Control Module that decides whether to
accept or reject the incoming slice requests. This decision helps the
INP manage its underlying resources efficiently, achieve fairness in
resource allocation among different slice types and increase its
revenue, while minimizing SLA violation among admitted slices.

In this work, we have used the Prioritised Experience Replay-based Deep
Q-Network with N-step return (N-PERDQN) Reinforcement learning (RL)
approach to solve the slice admission control and associated resource
allocation problem in a dynamic environment. We have also used Long
Short-Term Memory (LSTM) to predict the admitted elastic slices' future
resource requirements to share resources among elastic slices more
efficiently. This enables the INP to accept more slice requests to
increase its revenue while maximizing system utilization. The system has
been modeled using the SimPy discrete-event simulator and TensorFlow
machine learning. Performance metrics studied include INP revenue
generated, class-specific slice acceptance rate, and resource
utilization. The proposed N-PERDQN approach performs better than other
RL and heuristic methods by up to 10% on average in terms of generating
INP revenue. Also, N-PERDQN+LSTM accepts around 6% more slice requests
than N-PERDQN without LSTM. We have also tested the performance of the
proposed algorithm with an offline algorithm that has complete slice
request information. The proposed algorithm can achieve close to 82% to
86% in terms of revenue gain.