The rapid development of the Internet of Things (IoT) in recent years has resulted in insufficient Terminal Device (TD) computing capabilities [1]. To solve this issue, computing task offloading technology moves tasks to other servers for execution [2]. With the support of task offloading technology, TDs can delegate computing-intensive tasks to cloud computing platforms with sufficient resources, clout-lets in local area networks, and various edge computing nodes [3]. However, resource-rich cloud computing platforms are usually far away from TD [4], which will generate more network transmission delays [5]. In addition, traditional edge computing nodes and cloud-lets cannot support TD mobility well. Although Mobile Cloud Computing (MCC) supports the mobility of terminal devices, it still cannot solve the problem of high network communication latency [6].

The latest development trend is to offload the computing tasks of TDs to the Multi-access Edge Computing (MEC) network. Since the network communication distance between MEC and TDs is usually only one hop, and MEC supports the mobility of TDs very well [7]. However, MEC is different from MCC, and MEC has relatively limited computing resources. Therefore, task offloading decisions and resource allocation need to be optimized [8].

A large number of existing solutions to support MEC computing decision-making and resource allocation introduce traditional mathematical algorithms, heuristic algorithms, AI-based methods, etc. [9]. However, they only directly decide whether the computing task is run locally on the TD or on the MEC based on the decision result of the current time slot. Although some algorithms consider the problem of task offloading for multiple time slots in the future, they do not consider the problem of maintaining task offloading requests to strive for more offloading opportunities in multiple time slots in the future and notify TD as early as possible to start task execution locally. In MEC under conditions of relatively insufficient resources. This way, TDs may lose task offloading opportunities in the next few time slots and cause an increase in the overall task execution delay.

In this work, we focus on optimizing the task offloading decision optimization problem of IoT computing tasks under insufficient MEC resources. To overcome the limitations of single timeslot optimization, this study adopts a combined approach utilizing task offloading load prediction and virtual decision-making for subsequent timeslots to seek additional available resources for task offloading requests. In the proposed two-stage task offloading decision framework, the Deep Q-Network (DQN)-based decision generation algorithm and the Long Short-Term Memory (LSTM)-based task request time series prediction algorithm jointly predict the possibility of each offloading request being approved in future time slots, thereby determining the optimal offloading of the offloading task. strategies and reduce overall task execution time.

This study provides an in-depth analysis of the performance limitations of existing IoT task offloading methods in resource-constrained MEC networks and elucidates that the cause of this problem is the lack of ability to maximize the potential acceptance opportunities for offloading requests.

The rest of this study is organized as follows. Section 2 discussed the related work of the IoT task offloading method in MEC. Section 3 illustrated the proposed two-stage task offloading decision framework. Section 4 presented the performance evaluation and result discussion. The conclusion was given in Section 5.

As a key technique in IoT task offloading, the optimization of offloading decisions and resource allocation has been extensively studied in the past decades [10–13]. Classic task decision-making methods based on mathematical optimization are relatively mature. Such methods can usually find the global optimal solution in a strict mathematical sense [14–16]. However, this method requires mathematical modeling for specific application scenarios, so it has poor dynamic adaptability. In addition, this type of optimization method performs poorly when dealing with complex constrained problems with high-dimensional nonlinearity. The task offloading decision-making method based on game theory allows multiple participants to negotiate, and this method is suitable for adversarial task offloading and resource allocation scenarios [17,18]. This method is usually more conducive to ensuring that all parties involved receive relatively fair benefits. In addition, task offloading solutions based on game theory can better avoid system bottlenecks and single-point failures. However, such methods lead to a rapid increase in computational complexity when the number of participants is large, resulting in solution difficulties. In addition, this type of method may have difficulty converging in a dynamic system environment, and the game equilibrium state may be unstable. The task offloading optimization method based on fuzzy theory can more comprehensively consider multiple factors that affect decision-making effects and can use simpler models to describe complex real-life problems, thereby reducing the complexity of problem modeling. However, the performance of such methods is usually poor and the system output results are not intuitive. Based on heuristic methods, we do not seek the absolute optimal solution but seek relatively better solutions [19,20]. This method is cost-effective and suitable for solving large-scale problems. However, such methods are prone to falling into local optimal problems.

In order to cope with large-scale and dynamic complex computing task offloading application scenarios, various decision-making optimization methods based on AI technology have emerged in large numbers in the past decade [21,22]. This method has self-learning capabilities and can continuously optimize the model itself based on historical data and experience. In addition, AI-based task offloading decision-making methods can usually discover complex relationships hidden in high-dimensional data, so it is easy to find the global optimal solution. Moreover, this type of method can better adapt to dynamic changes in the network environment [23,24]. However, the methods discussed above lack the prediction of the offloading environment status for multiple time slots.

Besides, some researchers consider prediction with the computational load on the edge server. However, such studies mainly focus on the prediction of load on computing nodes but ignore the prediction of characteristics of task offloading requests from TD [25,26]. In addition, the AI models selected in a small number of studies can generate predictions of task scheduling decisions for multiple time slots in the future. However, no consideration is given to finding as many offloading opportunities as possible for offloading tasks in multiple time slots. We selected representative research works of different technical classifications for analysis, which cover common binary and partial task offloading optimization methods. The focus of the study is to analyze the start time of local execution tasks after the offload request is rejected. Table 1 shows the characteristics of representative task offloading algorithms to initiate local task execution.

The available resources of the MEC change dynamically due to the dynamics of task offloading requests and the completion of the running of computing tasks on the MEC. According to the review of literature, the majority of decision mechanisms only rely on the current slot’s task offloading network condition to make the final offloading decision. This results in rejected tasks immediately starting execution locally on the TD, missing the opportunity to seek edge computing resources in more time slots. We illustrate this issue with Fig. 1, which represents a common problem widely observed in similar studies [28,33,35,36].

It can be found in Fig. 1 that appropriately retrying to obtain the opportunity of task offloading in several adjacent decision cycles will be beneficial to reducing the completion time of the task after the first task offloading application is rejected. However, starting local task execution prematurely and excessive offload retries will increase task completion time.

In this section, the Two-Stage Task Offloading Decision Framework (TSODF) is proposed to allow offloading decision attempts across multiple time slots. In this method, best-effort decision optimization is performed on offloading requests and task allocation in each slot based on Deep Q-learning. In addition, the historical TD task offloading requests are input into the prediction model based on SLTM in time series to predict the task request load of multiple time slots in the future. Furthermore, the offloading decision for pending requests in future time slots is reasoned by combining the inferred MEC available computing resources and predicted TD task offloading requests. This enables the assessment of the cost and benefits of pending task requests. Subsequently, optimized decisions with higher request acceptance rates are output to reduce task execution latency and energy consumption.

We consider a MEC system, where M IoT TDs and an EMC server. Let D={d1,d2,…,dm}, denote the sets of the TDs. The MEC server S has powerful computing power Cs and unlimited battery life. We assume that every computational task is indivisible and let T={t1,t2,…,tn}. Each TD can choose to execute tasks locally or offload computing tasks to MEC for execution. Fig. 2 depicts the task offloading system model targeted by TSODF, referencing a typical MEC network architecture [37].

In the proposed framework, the DQN-based decision-making model performs best-effort optimized task offloading decisions for each time slot. Based on the allocated MEC server resource records, TSODF estimates the CPU resources available in each future time slot. The LSTM-based prediction module predicts task offloading requests in several future time slots based on historical TDs task offloading request records. TSODF jointly analyzes the above information to determine whether to continue to look for offloading opportunities for waiting task offloading requests, and outputs the optimized final decision result. Fig. 3 shows the proposed system framework of TSODF.

TSODF contains two main work stages, in which stage 1 performs the generation of optimal offloading decisions in each decision cycle based on deep reinforcement learning; stage 2 performs task offloading request load prediction based on LSTM. In addition, TSODF estimates the available resources of the future MEC server based on the decision results that have been generated. Thus, TSODF infers the probability that a task offloading request will be accepted in the future for offloading tasks that have not yet been accepted. Based on this mechanism, task offloading requests will be suspended waiting for an upcoming acceptance opportunity or starting local execution immediately.

The DQN-based but time-slot offloading decision generation method is discussed in Section 3.1. Section 3.2 introduces the task offloading request prediction method based on LSTM. The complete TSODF framework workflow is given in Section 3.3.

In this section, numerical results are presented to evaluate the performance of the proposed TSODF framework. Simulation environment information is given in Section 4.1. The results of the experiments and data analysis are discussed in Section 4.2.

In this work, the optimization of IoT task offloading decisions in MEC is studied in depth. Different from previous work, the proposed Two-Stage Task Offloading Decision Framework (TSODF) considers the continuous optimization of multiple attempts for unapproved offload requests when MEC resources are insufficient. The TSODF is designed to explore available edge computing network resources to accept task offloading requests from TDs. In this way, the success rate of task offloading and the utilization of edge computing resources are increased, resulting in overall reductions in task completion time and energy consumption. In addition, the request eviction reduces the delay for local task execution in waiting for decision-making. Simulation results show that the performance of the proposed task offloading framework improves the overall task offloading rate and reduces the overall task execution latency and energy consumption compared to the classical DQN-based task offloading decision scheme. Although only the case of DQN-based offloading decision generation techniques is considered in this paper, the proposed framework can be easily extended to the scenarios with the different types of offloading decision generation techniques. For future work, we will investigate optimization methods for IoT task offloading decisions in MEC environments supporting scalable IoT networks.