ABSTRACT
offloading scheduling goals: minimizing execution latency and maximizing user utility.
Specifically, the offloading scheduling decision process on tasks with DAG dependency
is modeled as a Markov decision process (MDP). A recurrent neural network-based se-
quence to sequence parameter-shared deep neural network architecture, as well as the
corresponding DAG embedding method, are proposed to approximate the MDP offload-
ing scheduling policy, which is then trained based on state-of-the-art proximal policy
optimization (PPO). The effectiveness and reliability of the proposed algorithm are ver-
ified through comparison with six baseline algorithms in different environments and for
different offloading scheduling goals.
Furthermore, also from the perspective of users, the offloading scheduling decision-
making problem in a highly dynamic vehicular MEC environment is studied, in which
all the dynamic factors, such as task arrival, task attribute, wireless channel, and user
mobility, are taken into account. This stochastic optimization problem is very difficult
for conventional solutions because of the sophisticated environment dynamics and vast
state space. A DRL-based dynamic offloading scheduling algorithm is designed, which
jointly solves “where” to schedule and “when” to schedule each task in the complex task
queue, to achieve the optimal long-term tradeoff between task latency and energy con-
sumption in such a complicated environment. A series of methods are adopted to improve
the training efficiency and convergence performance of the algorithm, including utilizing
PPO to ensure the efficiency and stability of the training process, embedding a convolu-
tional neural network into the policy network to extract the key features of the complex
task queue, and adjusting the state and reward of DRL to avoid inefficient exploration in
the training process. Extensive simulation experiments demonstrate that the performance
of the proposed algorithm is much higher than that of traditional baseline algorithms in
different environments and user preferences.
Finally, from the perspective of operators, the admission decision and resource al-
location strategy among multiple moving users in a vehicular MEC environment is in-
vestigated, in which the constraints of computation and communication resource limi-
tation, task deadline requirements, and the mobility of users are taken into account, in
order to maximize the overall system utility. The optimization problem is formulated as
a mixed-integer non-linear programming (MINLP), and a heuristic multi-user mobility-
aware offloading decision algorithm with polynomial time complexity is proposed. The
original global optimization problem is converted into a finite number of local optimiza-
IV
