Delay Constrained Resource Allocation for NOMA Enabled Satellite Internet of Things with Deep Reinforcement Learning

With the ever increasing requirement of transferring
data from/to smart users within a wide area, satellite internet of
things (S-IoT) networks has emerged as a promising paradigm
to provide cost-effective solution for remote and disaster areas.
Taking into account the diverse link qualities and delay qualityof-
service (QoS) requirements of S-IoT devices, we introduce a
power domain non-orthogonal multiple access (NOMA) scheme
in the downlink S-IoT networks to enhance resource utilization
efficiency and employ the concept of effective capacity
to show delay-QoS requirements of S-IoT traffics. Firstly, resource
allocation among NOMA users is formulated with the
aim of maximizing sum effective capacity of the S-IoT while
meeting the minimum capacity constraint of each user. Due to
the intractability and non-convexity of the initial optimization
problem, especially in the case of large-scale user-pair in NOMA
enabled S-IoT. This paper employs a deep reinforcement learning
(DRL) algorithm for dynamic resource allocation. Specifically,
channel conditions and/or delay-QoS requirements of NOMA
users are carefully selected as state according to exact closed-form
expressions as well as low-SNR and high-SNR approximations,
a deep Q network is first adopted to yet reward and output
the optimum power allocation coefficients for all users, and then
learn to adjust the allocation policy by updating the weights
of neural networks using gained experiences. Simulation results
are provided to demonstrate that with a proper discount factor,
reward design, and training mechanism, the proposed DRL
based power allocation scheme can output optimal/near-optimal
action in each time slot, and thus, provide superior performance
than that achieved with a fixed power allocation strategy and
orthogonal multiple access (OMA) scheme.