The digital twin (DT) assisted mobile edge computing (MEC), which adopts DT to bridge cyber and physical systems by generating digital replicas of real entities, is proposed as an effective solution to manage the proliferating Internet of Things (IoT) networks. However, it is challenging to continuously maintain the digital representation of physical system because of the limited battery and computation capabilities of IoT devices. In this paper, we present a new paradigm of DT assisted wireless powered MEC (WP-MEC) network, where IoT devices can harvest energy from hybrid access points (HAPs) and upload sensing data to maintain DT. The DT, in turn, is responsible for efficiently making offloading decisions and resource allocations to maximize the computing data volume of the WP-MEC. To maximize the computation rate and maintain the DT assisted WP-MEC network, we formulate a mixed-integer nonlinear programming (MINLP) problem that addresses task offloading, wireless power transfer (WPT) duration, energy allocation, DT uploading duration, and data processing duration. Initially, we simplify the original problem through mathematical derivation. Subsequently, we design a mixed deep reinforcement learning (mixed-DRL) framework comprising two deep neural networks (DNNs) that output discrete offloading decisions and continuous resource allocations to achieve a higher volume of processed task data. Extensive experiments demonstrate that the proposed mixed-DRL framework with the self-designed DRL algorithms can improve the processed data volume by 30% compared to the advanced algorithm in the literature.

In the ongoing 5G and upcoming 6G eras, the intelligent Internet of Things (IoT) network will take increasingly important responsibility for industrial production, daily life and so on. The IoT devices with limited battery size and computing ability cannot meet many applications brought out by the data-driven artificial intelligence technique. The combination of wireless power transfer (WPT) and edge computing is regarded as an effective solution to this dilemma. IoT devices can collect radio frequency energy provided by hybrid access points (HAPs) to process data locally or offload data to the edge servers of HAPs. However, how to efficiently make offloading decisions and allocate resource is challenging, especially for the networks with multiple HAPs. In this paper, we consider the sum computation rate maximization problem for a WPT empowered IoT network with multiple HAPs and IoT devices. The problem is formulated as a mixed-integer nonlinear programming problem. To solve this problem efficiently, we decompose it into a top-problem of optimizing offloading decisions and a sub-problem of optimizing time allocation under the given offloading decisions. We propose a deep reinforcement learning (DRL) based algorithm to output the near-optimal offloading decision and design an efficient algorithm based on Lagrangian duality method to obtain the consequent optimal time allocation. Simulations verified that the proposed DRL-based algorithm can achieve more than 95 percent of the maximal computation rate with low complexity. Compared with the common actor-critic algorithm, the proposed algorithm has the substantial advantage in convergence speed, achieved computation rate and running time.

In recent years, wireless powered mobile edge computing (WP-MEC) is one of solutions to the problem of insufficient computing power and battery capacity of current edge devices (EDs). In this paper, we consider a WP-MEC network with multiple EDs and study the problem of maximizing the long-term computation rate of the system under the premise of maintaining the stability of the system data queue. Specifically, the objective function is described as a complex non-convex problem, we used Lyapunov optimization theory to decouple the multi-stage continuous random problem into sub-problems of deterministic frames. The problem of determining frames requires joint optimization of wireless power transfer (WPT) duration, local computing frequency, transmission duration, and energy required for offloading. In order to efficiently optimize these variables, we design a DRL-based algorithm combined with the CVX solver, the DRL algorithm learns the WPT duration, and the convex optimization algorithm obtains the system offloading strategy. From the simulation results, our algorithm can achieve performance close to that of the one-dimension exhaustive search algorithm while ensuring the stability of the data queue.