Sage Journals: Discover world-class research

Abstract

Mobile micro-learning has received extensive attention in the research of smart cities because it is a novel fusion service mode of the mobile Internet, cloud computing, and micro-learning. However, due to the explosively increased applications of the mobile micro-learning and the limited resources of mobile terminals, an effective energy saving approach for mobile micro-learning is urgently required. For this end, this article proposes an efficient task joint execution strategy to reduce energy consumption. First, a new matching method of time series is proposed to obtain the latest requested record, which can provide guidance for the selection of a future service mode. Second, a mapping-level service mode and a cloud-level service mode are proposed to achieve seamless switching. Finally, the genetic algorithm is used to find the optimal executive strategy. In addition, the experimental results show that the proposed method can effectively realize the target of energy saving by using real data set.

Keywords

Green city mobile micro-learning Cloudlet time-series matching genetic algorithm

Introduction

As a novel fusion service mode of the mobile Internet, cloud computing, and micro-learning, mobile micro-learning has attracted extensive attention in the research of smart cities. Moreover, its running is not limited by time, space, and region; the mobile micro-learning business presents an explosive growth trend, which needs more computing, storage, bandwidth, and energy. However, since the mobile terminal takes a great deal of power to deal with the applications of mobile micro learning, which directly restricts the development of mobile micro-learning.. For this reason, we focus on the efficient, reliable, and low energy consumption service in mobile micro-learning.

The “mobile-Cloudlet-cloud” architecture is generated by the fusion of mobile computing and cloud computing technology. As the middle tier of the architecture, Cloudlet can bring the cloud closer to the terminal user. Therefore, it can avoid some problems, for example, the latency and bandwidth limitation of Wide Area Network. In addition, Cloudlet is a mobility-enhanced small-scale cloud data center, which is placed at the edge of the Internet. Thus, it also has the advantage of the cloud platform, such as powerful processing power and storage capacity. Furthermore, to reduce computation and storage pressure of mobile terminals, offloading compute-intensive task to the remote resource-rich cloud server has been proven to be an effective method.^1–4 However, the distance between the physical locations of cloud servers and the mobile terminal users is remote, which could cause network latency and increase battery-empowered mobile terminal resource consumption. Therefore, this article uses Cloudlet to replace cloud as a service provider.

Introducing Cloudlet into the micro-learning is a promising method for energy saving because users can make full use of high computation and huge storage power of Cloudlet. However, many researchers focus mainly on the user behavior collection,⁵ learning style analysis,⁶ framework design,⁷ high attrition rates,⁸ and so on. The energy consumption of mobile micro-learning is rarely discussed.

However, low energy consumption and high dependability are two crucial factors for the development of mobile micro-learning. For this reason, we propose a task joint execution strategy to reduce mobile terminal energy consumption and ensure the dependability of mobile micro-learning within the threshold time. The major contributions of this article can be summarized as follows.

First, we propose a time-series framework to describe the process of mobile micro-learning. Meanwhile, we also propose a new time-series matching method to obtain user history request record. In addition, our method overcomes the limitations of traditional studies, for example, splitting internal similarity between the user requests and using fixed resource consumption.

Second, we propose a joint task execution strategy that can achieve seamless switching between mapping-level service mode and cloud-level service mode based on users’ personalized requirement and energy-aware.

Third, in order to ensure green and reliability of mobile micro-learning, we propose a new algorithm based on time-series matching method and genetic algorithm (GA). The simulation results show the superiority of our proposed algorithm.

The remaining article is organized as follows. We review related literature in the “Related work” section. We briefly describe mobile micro-learning system architecture and propose a task joint execution strategy model in the “Task joint execution strategy modeling” section. We formulate energy consumption based on our model in the “Problem formulations” section. We propose a new algorithm based on the model in the “Optimization algorithm based on GA” section. The simulation and numerical results are presented in the “Simulation and analysis” section. Finally, the conclusion and the future works are provided in the “Conclusions and future works” section.

Related work

Because of the convenience and ubiquity of cloud services, people are becoming more dependent on smart cities. Facing irreversible development tendency of smart cities, scholars devote great effort to research-related projects, such as energy, construction, transportation, healthcare, education, human services, and so on.^9–12 However, smart learning and education are the key areas of the smart city in those projects.¹³ At the same time, a new round of smart city construction has also put forward new requirements for education. Thus, mobile micro-learning, as a kind of intelligent education, has received extensive attention in the smart city. For example, Markovic and Sofronijevic¹⁴ focus on large-scale learning resources in the smart city and collect education-related data. Salerno¹⁵ mainly studies relationships between knowledge society and smart city. The knowledge society allows the smart city to use potential resources for the benefit of their citizens, and the smart city offers live, advanced, first-quality knowledge resources to the knowledge society. A real application of intelligent education is established in Liu,¹⁶ which is English teaching platform for the non-English majors in colleges. Malek et al.¹⁷ design a framework for smart city learning scenarios. The framework has friendly graphical interfaces, and it simulates adaptive learning activities of learners. The customized smart teaching and smart learning system are proposed in Elhoseny et al.¹⁸ Hammad and Ludlow¹⁹ propose a structural framework for the successful application of smart learning environments in the context of smart city governance. A real-life case “Smart People for Smart Cities!” is studied in Kadar,²⁰ which simulates the real conditions of the smart city and creates an enhanced learning and teaching environment. Meanwhile, Yin et al.²¹ present context-aware mobile learning for professional situations in the smart city. In addition, Souza and Amaral²² study the theory, concept, and model of mobile micro-learning. Sun et al.²³ design micro learning as a service (MLaaS) with the help of cloud platform. The MLaaS can provide adaptive micro-learning contents and institute the learning paths customized for each individual learner. Kim et al.²⁴ discuss the impact of mobile wireless technology on mobile learning. Chen et al.²⁵ collect and analyze a large number of non-organized micro-learning units and excavate users’ personalized needs. To the best of our knowledge, the energy consumption of mobile micro-learning is rarely discussed.

The research on energy saving of mobile terminals has been a hot topic in recent years. Offloading compute-intensive task to resource-rich cloud server has been considered as a potential energy saving solution for the mobile terminal. Fekete et al.²⁶ propose an energy-efficient job scheduling model and measurement architecture. Shu et al.²⁷ study energy-efficient data transmission strategy between cloud and mobile terminal based on Lyapunov optimization. Haridas et al.²⁸ propose a cloud offloading algorithm under the guidance of inherent patterns of task sequences. Chatzopoulos et al.²⁹ propose a framework that integrates incentive scheme and reputation mechanism, which is a Device-to-Device offloading strategy.

During the process of task offloading, the location of the task execution after the offloading is a difficult problem. Mobile Edge Computing (MEC) provides information technology (IT) and cloud computing capabilities within the Radio Access Network (RAN) in close proximity to mobile subscribers.³⁰ Thus, the edge server of the cloud data center is considered to be a good task execution location. Intharawijitr et al.³¹ study communication and computation delay of the mobile application by introducing MEC; the results show that relaxing latency constraints and choosing a server with minimal total latency can improve the system performance significantly. Wei et al.³² propose “MVR,” which is an innovative architecture for computation offloading in MEC. To minimize service delay and operational cost, Xu et al.³³ put forward an efficient reinforcement learning based on resource management algorithm. The algorithm can learn the optimal policy of dynamic workload offloading and edge server provisioning. Wang et al.³⁴ present an integrated framework based on computation offloading and interference management in wireless cellular networks. The framework can make decisions, such as computation offloading, resource allocation and optimal, by the trade-off cost between the local computing and offloading computing. Cao and Cai³⁵ discuss computation offloading based on game theory in the multi-channel wireless competition environment. A new scheduling technique is proposed by Abd et al.,³⁶ which can achieve energy efficiency, reliability, and time management at the same time. The main purpose of Cloudlet is to support resource-intensive and interactive mobile applications by providing powerful computing resources to mobile devices with lower latency. Therefore, this article uses Cloudlet assistance to model for the energy consumption in the mobile micro-learning from different angles.

The user behavior of mobile micro-learning is close to the lifestyle, job category, operational habit, interest, and so on. Although user request has randomness and suddenness to a certain extent, user operation behavior is relatively stable in the long term. Lippi et al.³⁷ emphasize the correlation among time series. Ye et al.³⁸ point out that the user requests are time series; thus, the problem of cloud service composition can be solved through time-series similarity. Wang et al.³⁹ study access control based on resource allocation. The researches provide ideas and methods for the analysis of user requests. Therefore, we can take positive efforts to analyze the changes of user demand and establish the performance model of mobile micro-learning according to the similarity of the time series.

Task joint execution strategy modeling

At the beginning of this section, we briefly describe the system architecture of mobile micro-learning. Then, we propose a new model of mobile micro-learning based on time-series similarity and GA.

Mobile micro-learning system architecture

Mobile micro-learning is the combination of mobile learning and micro-learning. As an ideal non-traditional learning mode, its mobility can meet the needs of learners in a dynamic environment. The micro-characteristic can facilitate learners to learn in a fragmented time. The ubiquity and interactivity mean that learning happens anytime, anywhere, and on demand. Therefore, it can be defined as a learning model. In this model, learners use multiple devices to access Cloudlet platform through the mobile network and obtain the learning resources from Cloudlet platform whenever and wherever they need.

The system architecture of the mobile micro-learning is shown in Figure 1. As shown above, mobile micro-learning is influenced by the mobile terminal, mobile network, and learning resources of the cloud platform. More specifically, the platform function of the mobile terminal determines whether the mobile micro-learning process can be completed or not. Mobile network status affects the dependability of operation selection. The interactivity and accuracy of learning content are the two main factors that affect mobile micro-learning effects. In order to provide users with continuous cloud services and achieve the target of energy saving, we will consider mobile micro-learning from three aspects that are the resource of mobile terminal, the status of mobile network, and the service resource of cloud platform.

Figure 1.

Mobile micro-learning system architecture.

From the above description, we know that mobile micro-learning is tightly integrated with the smart city. On one hand, smart city has a wealth of learning resources and advanced infrastructure. The development of smart city will promote fusion between mobile micro-learning data and smart city data, and achieve unified management of mobile micro-learning data. Therefore, learners can learn without the limited time and space. On the other hand, mobile micro-learning should not be a self-enclosed area. It is necessary to make mobile micro-learning combined with the data in smart city, because it can provide the requisite condition for seamless learning.

In Figure 1, $S = {s_{1}, s_{2}, \dots, s_{k}}$ is service set, where $s_{k}$ denotes the $k th$ service. Because of the dynamic of the network environment and the personalization of user requests, the content of service may be updated at different time, but the service is still the same type. Therefore, we classify these services. For example, $s_{k - 1}$ and $s_{k}$ are the same type of services; we define a set as $Set_o = {s_{k - 1}, s_{k}}$ , where $Set_o$ denotes the $o th$ category of service.

Task joint execution strategy model

The task joint execution strategy model is shown in Figure 2. The model consists of three parts that are user module, service mode selection module, and GA optimization module. Its main function is to select the optimal strategy of task execution that can achieve energy saving and high dependability within the threshold time. The major contents of this model are explained in detail in the next paragraph.

Figure 2.

Task joint execution strategy model for mobile micro-learning.

Users’ personalized demands are obtained by the user module, which will provide guidance for service mode selection. As for service mode selection module, the initial service mode selection is the top priority according to users’ personalized demand and time-series matching method. For example, if a user requires the cloud-level service mode to provide service, the system will directly put the task into cloud-level service mode database, and vice versa. Else, we choose the suitable service mode under the guidance of time-series matching method. The implementation process of time-series matching method will be discussed in detail in the “Time series matching method” section. In service mode selection module, mapping-level service mode and cloud-level service mode are proposed to achieve seamless switching. Its working process is presented next. First, the time-series matching method is adopted to get the latest request record, which is consistent with the requested service. If the record has existed, we can get its resource consumption. If not, we directly put the task into cloud-level service mode database. Second, we can use the control monitor to obtain the current network resource situation. If the existing resource can meet users’ resource needs and cannot affect other assemblies on the system, the service will be provided by mobile terminal mapping-level service mode. Otherwise, the task will be put into cloud-level service mode database. With all the operations above, we can get two databases: cloud-level service mode database and mapping-level service mode database. The main function of the GA optimization module is to put the request back together and try to find out the optimal task execution strategy, which is a task joint execution strategy that can achieve energy saving and reliability.

Problem formulations

In this section, we mainly discuss the time-series matching method and formulate energy consumption of mobile micro-learning according to the task joint execution strategy.

Transforming mobile micro-learning process into time-series model

To demonstrate the model, we define user set as $U = {u_{1}, u_{2}, \dots, u_{i}}$ , where $u_{i}$ denotes the $i th$ user. The $T = {t_{1}, t_{2}, \dots, t_{j}}$ denotes the time set, where $t_{j}$ is the $j th$ time slice. Thus, we define user request trajectory of user $i$ by $Q_{i} = {q_{i}^{1, 1}, q_{i}^{2, 2}, \dots, q_{i}^{j, k}}$ , where $q_{i}^{j, k}$ is a user request when user $i$ requests service $k$ at time $j$ . Because tasks can be associated with a request, a task set is created for user $i$ , which is formally defined as $Tas k_{i} = {{task}_{i}^{1, 1}, {task}_{i}^{1, 2}, \dots, {task}_{i}^{j, n}}$ , where $n$ is the $n th$ task. The ${task}_{i}^{j, n}$ is equivalent to $q_{i}^{j, k}$ for convenience in the next sections.

As we all know, the user request trajectory consisted of the long interest model and the short interest model. The short interest model records the user recent interests. If users care about something continuously during a period of time, it will be added to the long interest model. Therefore, the long interest model is made up of the short interest in different time slices. Because the short interest is a short-term burst of users’ interest, the trend of the long interest model will decline gradually. As a result of different background, interest, and motivation, everyone’s demand for learning resource is specific. However, the study reveals that personality has some correlation with users’ information source preferences. Therefore, regardless of which model you use, reviewing the history of user interest development and analyzing its evolution trends will provide guidance for studying users’ personalized demand to a certain extent.

Time-series matching method

In this section, we propose a time-series matching method. Our goal is to find the latest user history request exactly that is consistent with user current request. More importantly, the study results provide not only guidance on resources for the current user request but also reference for service mode selection of mobile micro-learning.

The time-series matching method is shown in Figure 3. As is shown, we define a five-step procedure to implement the time-series matching method. Before we discuss the major operations in detail, the degree of correlation between user requests is introduced to obtain relationships between one user request and another; it is

Sc = \frac{P_{sh} * M_{sh}}{T_{sh}} + \frac{P_{lo} * M_{lo}}{T_{lo}}

(1)

where $M_{sh}$ and $M_{lo}$ denote the number of successful matches that are obtained by the short interest model and the long interest model, respectively; $P_{sh}$ and $P_{lo}$ denote the scale weights of the short interest model and the long interest model, respectively; and $T_{sh}$ and $T_{lo}$ denote the total number of matches that are obtained by the short interest model and the long interest model, respectively. Next, we will introduce the process of time-series matching method in detail. First, for a user request $q_{\tilde{i}}^{\tilde{j}, \tilde{k}}$ , this method will find its corresponding historical trajectory in the historical database; we suppose the historical trajectory is $Q_{\tilde{i}}$ . Second, the time-series matching process is used to find $proAp$ that is a possible matching history request record. Third, we can get a candidate request set that is consisted of $proAp$ . Fourth, this method will calculate $Sc$ and sorted user request by order of $Sc$ . The larger the $Sc$ of a user request is, the higher the degree of correlation is. Finally, because the request must match the requested service by the user, exact character matching method is used to find a user request record, which is consistent with the current request. From the above steps, when the user requires this service again, if the service user requested is existent and never updated, the mapping-level service mode will provide the service for the user. Otherwise, the cloud-level service mode will provide service for the user. Hence, the time-series matching method provides guidance for the initial service mode selection. This method involves one important process, that is, time-series matching. The details are described as follows.

Figure 3.

Time-series matching method.

In the process of time-series matching, the historical trajectory is seen as a string, and the current user request is considered to be a substring. 1. we set three variables $σ$ , $κ$ , and $τ$ . $σ$ is the length of the match, which is a condition and uses to measure the similarity between two strings. For example, there are two strings that are “abc” and “bcd”. Because they have two same characters that are “b” and “c”, the match length is 2. $κ$ is match step that is equal to the length of the user request. Its function is to control the length of the substring moving forward. $τ$ is the number of successful matches, which is designed to distinguish the long interest models and the short interest models. If the number of successful matches between user request and historical trajectory is less than $τ$ , the historical trajectory is the long interest model. Otherwise, it is the short interest model. 2. the times series method will go through the historical request records in order from right to left until it finds the one that matches the current request. For example, if a user request is “abc,” the historical trajectory is “abdbbbbacabceab.” Therefore, “abdbbbbacabceab” is divided into “abd/bbb/bac/abc/eab” according to $κ = 3$ . 3. the times series method will compute the match length among “abc” and “eab”, “abc”, “bac”, “bbb”, “abd”. The match lengths of this example are 2, 3, 3, 1, and 2, respectively. 4, we judge what kind of history requests can become a possibility for successful historical request records according to $σ$ . If the length of succeeding matches between the user request and historical trajectory is greater than $σ$ , the historical trajectory may be a successful request record that is in the same with the current user request. 5. we can get a candidate request set.

When steps 1–5 has been completed, we can get the matching accuracy ratio; it is defined as

I_{AR} = \frac{A}{G}

(2)

where $A$ refers to the number of successfully matched request and $G$ is the total number of requests.

Utility function

In this section, we will discuss the energy consumption of two service modes, respectively; our motivation is to formulate the energy consumption of mobile micro-learning and provide guidance for GA optimal model.

We denote the Cloudlet set by $C_{let} = {c_{1}, c_{2}, \dots, c_{m}}$ , where $m$ represents the number of Cloudlet. Task size (in bits) of $u_{i}$ is $B_{size} = (b_{(i, 1)}^{1}, b_{(i, 2)}^{2}, b_{(i, 3)}^{3}, \dots, b_{(i, j)}^{k})$ , where $k$ represents the $k th$ task. The execution rate is $R_{cl}$ when the task is executed by Cloudlet and it is $R_{mp}$ when the task is executed by mobile terminal. In this article, we set $R_{cl}$ and $R_{mp}$ as constant values. The data transmission rate is defined as $R_{off} = [r_{1}, r_{2}, \dots, r_{m}]$ . At the same time, mobile terminal execution cost denotes as $P_{mp}^{power}$ , cloud execution (CE) cost is $P_{cl}^{power}$ , and offloading cost is $P_{off}^{power}$ . Then we can get $\tilde{k} th$ task execution cost of $\tilde{i} th$ user at $\tilde{j} th$ time. Next, we can formulate the time cost and the energy consumption in different service modes, respectively.

We assume that the execution time of the task is only influenced by the size of the task and the execution rate of the task. When the task is performed by the mapping-level service mode, the time cost is given by

T_{map - level}^{(\tilde{i}, \tilde{j}, \tilde{k})} = \frac{b_{(\tilde{i}, \tilde{j})}^{\tilde{k}}}{Rmp}

(3)

When the mapping-level service mode performs the task, the energy consumption cost is calculated as follows

E_{map - level}^{(\tilde{i}, \tilde{j}, \tilde{k})} = T_{mp}^{(\tilde{i}, \tilde{j}, \tilde{k})} * P_{mp}^{power}

(4)

Similar to equations (3) and (4), when the task is executed by Cloudlet platform, the time cost is

T_{cl}^{(\tilde{i}, \tilde{j}, \tilde{k})} = \frac{b_{(\tilde{i}, \tilde{j})}^{\tilde{k}}}{R_{cl}}

(5)

When Cloudlet platform performs the task, the energy consumption cost is calculated by equation (6)

E_{cl}^{(\tilde{i}, \tilde{k})} = T_{cl}^{(\tilde{i}, \tilde{j}, \tilde{k})} * P_{cl}^{power}

(6)

If the resource demand of a task is exceeding the available resources of mobile terminal, the mapping-level service mode cannot provide the required service to the mobile micro-learning users within the threshold time. Therefore, to ensure quality of service (QoS) of mobile micro-learning, the task is offloaded from the mapping-level service mode onto the cloud-level service mode. Facing with offloading process, we need to consider the cost of offloading.

The time cost of offloading process can be defined as follows

T_{off}^{(\tilde{i}, \tilde{j}, \tilde{k})} = \frac{b_{(\tilde{i}, \tilde{j})}^{\tilde{k}}}{R_{off}}

(7)

The energy cost of the offloading process is given by equation (8)

E_{off}^{(\tilde{i}, \tilde{j}, \tilde{k})} = T_{off}^{(\tilde{i}, \tilde{j}, \tilde{k})} * P_{off}^{power}

(8)

In this article, the cloud-level service mode consisted of the offloading process and execution process of task. Therefore, when the user obtains the service from the cloud-level service mode, the time cost is

T_{clo - level}^{(\tilde{i}, \tilde{j}, \tilde{k})} = T_{cl}^{(\tilde{i}, \tilde{j}, \tilde{k})} + T_{off}^{(\tilde{i}, \tilde{j}, \tilde{k})}

(9)

When cloud-level service mode provides the service, the energy cost can be formulated as

E_{clo - level}^{(\tilde{i}, \tilde{j}, \tilde{k})} = E_{cl}^{(\tilde{i}, \tilde{j}, \tilde{k})} + E_{off}^{(\tilde{i}, \tilde{j}, \tilde{k})}

(10)

The total time cost and energy cost are associated with the time cost and energy cost of each task. Therefore, when the service is provided by mapping-level service mode, we get the total energy cost

T_{map - level}^{total} = \sum_{\tilde{k} = 1}^{k} \sum_{\tilde{j} = 1}^{j} \sum_{\tilde{i} = 1}^{i} P_{mp}^{(\tilde{i}, \tilde{j}, \tilde{k})} * T_{map - level}^{(\tilde{i}, \tilde{j}, \tilde{k})}

(11)

where $P_{mp}^{(\tilde{i}, \tilde{j}, \tilde{k})}$ is probability that the task $\tilde{k}$ of user $\tilde{i}$ is executed at time $\tilde{j}$ .

When the service is provided by mapping-level service mode, the total energy cost can be expressed as

E_{map - level}^{total} = \sum_{\tilde{k} = 1}^{k} \sum_{\tilde{j} = 1}^{j} \sum_{\tilde{i} = 1}^{i} P_{mp}^{(\tilde{i}, \tilde{j}, \tilde{k})} * E_{map - level}^{(\tilde{i}, \tilde{j}, \tilde{k})}

(12)

When users get service from cloud-level service mode, the total time cost is defined in equation (13)

T_{clo - level}^{total} = \sum_{\tilde{k} = 1}^{k} \sum_{\tilde{j} = 1}^{j} \sum_{\tilde{i} = 1}^{i} P_{cl}^{(\tilde{i}, \tilde{j}, \tilde{k})} * T_{clo - level}^{(\tilde{i}, \tilde{j}, \tilde{k})}

(13)

When users get service from cloud-level service mode, the total energy consumption cost is given as follows

E_{clo - level}^{total} = \sum_{\tilde{k} = 1}^{k} \sum_{\tilde{j} = 1}^{j} \sum_{\tilde{i} = 1}^{i} P_{cl}^{(\tilde{i}, \tilde{j}, \tilde{k})} * E_{clo - level}^{(\tilde{i}, \tilde{j}, \tilde{k})}

(14)

The mobile micro-learning may be affected when Cloudlet capacity is insufficient or mobile terminal is beyond the service scope of Cloudlet. To ensure the smooth completion of mobile micro-learning and achieve the target of energy saving, we need to find a suitable service mode and offloading destination. For this purpose, we assume the demand for central processing unit (CPU), memory, and bandwidth is $d_{(\tilde{i}, \tilde{j}, \tilde{k})}^{cpu}$ , $d_{(\tilde{i}, \tilde{j}, \tilde{k})}^{mem}$ , and $d_{(\tilde{i}, \tilde{j}, \tilde{k})}^{bw}$ , respectively, when task $\tilde{k}$ of user $\tilde{i}$ is executed at time $\tilde{j}$ . We define the binary function $P_{(\tilde{i}, \tilde{j}, \tilde{k})}^{\tilde{m}}$ for the task $\tilde{k}$ , where $P_{(\tilde{i}, \tilde{j}, \tilde{k})}^{\tilde{m}} = 1$ if $\tilde{k} th$ task of user $\tilde{i}$ is executed on the Cloudlet $\tilde{m}$ at time $\tilde{j}$ , and otherwise $P_{(\tilde{i}, \tilde{j}, \tilde{k})}^{\tilde{m}} = 0$ . After a period of time, the available resources of system are $C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{cpu}$ , $C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{mem}$ , and $C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{bw}$ , respectively, where $C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{cpu}$ denotes available CPU resource, $C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{mem}$ denotes available memory resource, and $C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{bw}$ denotes available bandwidth resource. The $C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{cpu}$ , $C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{mem}$ , and $C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{bw}$ are formulated as follows

C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{cpu} = 1 - \sum_{\tilde{k} = 1}^{k - 1} \sum_{\tilde{j} = 1}^{j - 1} \sum_{\tilde{i} = 1}^{i - 1} P_{(\tilde{i}, \tilde{j}, \tilde{k})}^{\tilde{m}} * d_{(\tilde{i}, \tilde{j}, \tilde{k})}^{cpu}

(15)

C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{mem} = 1 - \sum_{\tilde{k} = 1}^{k - 1} \sum_{\tilde{j} = 1}^{j - 1} \sum_{\tilde{i} = 1}^{i - 1} P_{(\tilde{i}, \tilde{j}, \tilde{k})}^{\tilde{m}} * d_{(\tilde{i}, \tilde{j}, \tilde{k})}^{mem}

(16)

C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{bw} = 1 - \sum_{\tilde{k} = 1}^{k - 1} \sum_{\tilde{j} = 1}^{j - 1} \sum_{\tilde{i} = 1}^{i - 1} P_{(\tilde{i}, \tilde{j}, \tilde{k})}^{\tilde{m}} * d_{(\tilde{i}, \tilde{j}, \tilde{k})}^{bw}

(17)

A normalization process is necessary while CPU, memory, and bandwidth have different units. Next, we illustrate it in the content of CPU but is easily generalized to memory and bandwidth

∥ C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{c p u} ∥ = {\begin{matrix} \frac{{(C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{c p u})}^{m a x} - C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{c p u}}{{(C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{c p u})}^{m a x} - {(C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{c p u})}^{m i n}} & \begin{matrix} {(C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{c p u})}^{m a x} \\ \neq {(C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{c p u})}^{m i n} \end{matrix} \\ 1 & e l s e \end{matrix}

(18)

where $(C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{cpu})^{\max}$ is the maximum CPU consumption of the Cloudlet $\tilde{m}$ . $(C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{cpu})^{\min}$ denotes the minimum CPU consumption of the Cloudlet.

We extend the normalization process to memory and bandwidth. Similar to equation (18), we defined a normalization process of memory and bandwidth that is shown in equations (19) and (20), respectively

∥ C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{m e m} ∥ = {\begin{matrix} \frac{{(C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{m e m})}^{m a x} - C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{m e m}}{{(C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{m e m})}^{m a x} - {(C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{m e m})}^{m i n}} & \begin{matrix} {(C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{m e m})}^{m a x} \\ \neq {(C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{m e m})}^{m i n} \end{matrix} \\ 1 & e l s e \end{matrix}

(19)

where $(C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{mem})^{\max}$ is the maximum memory consumption of the Cloudlet $\tilde{m}$ . $(C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{mem})^{\min}$ is the minimum memory consumption of the Cloudlet $\tilde{m}$

∥ C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{b w} ∥ = {\begin{matrix} \frac{{(C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{b w})}^{m a x} - C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{b w}}{{(C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{b w})}^{m a x} - {(C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{b w})}^{m i n}} & \begin{matrix} {(C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{b w})}^{m a x} \\ \neq {(C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{b w})}^{m i n} \end{matrix} \\ 1 & e l s e \end{matrix}

(20)

where $(C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{bw})^{\max}$ is the maximum bandwidth consumption of the Cloudlet $\tilde{m}$ . $(C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{bw})^{\min}$ is the minimum bandwidth consumption of the Cloudlet $\tilde{m}$ .

There is no clear definition of resource consumption in a dynamic network environment. Therefore, we cannot accurately calculate the resource consumption of each task. Similar to previous work,⁴⁰ we give a definition for the available resource of $\tilde{m} th$ Cloudlet at time $\tilde{j}$ as follows

\begin{array}{l} C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})} = α ∥ C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{c p u} ∥ \\ + β ∥ C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{m e m} ∥ + γ ∥ C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{b w} ∥ \end{array}

(21)

where $α, β, γ \in [0, 1]$ are weights of the available resources of the system during the task execution after normalization process. Specifically, $α$ is weight of available CPU resource, $β$ is weight of available memory resource, and $γ$ is weight of available bandwidth resource. In this article, we only consider CPU, memory, and bandwidth resources. Therefore, $α$ , $β$ , and $γ$ satisfy the normalization relationship $α + β + γ = 1$ . In general, the higher the $C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}$ is, the better the performance of the service mode is.

The insufficiency of any kind of resource can be a bottleneck, which influences the performance of the system. Therefore, to guarantee the dependability of service, the task success ratio as an evaluation index is defined as

I_{OSR} = \frac{O}{V}

(22)

where $O$ denotes the number of tasks that are executed successfully. $V$ is the total number of tasks. The larger the task success ratio is, the optimal the strategy is.

In order to meet the goal of the green city, we need to find the minimum energy consumption cost. Meanwhile, in order to ensure user satisfaction, response time of user request does not exceed the time threshold that is expressed by $T_{th}$ . Therefore, by combining the utility function and system constraints, we get

F (E) = \min {E_{map - level}^{total} + E_{clo - level}^{total}}

(23)

F (I) = \max {I_{OSR}}

(24)

Subject to

\begin{matrix} \forall u_{\tilde{i}} \in U, t_{\tilde{j}} \in T, c_{\tilde{m}} \in C_{let, s_{\tilde{k}}} \in Set_o \\ \forall \tilde{i}, \tilde{j}, \tilde{m}, \tilde{k} \in N \end{matrix}

(25)

0 \leq d_{(\tilde{i}, \tilde{j}, \tilde{k})}^{cpu} < C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{cpu}

(26)

0 \leq d_{(\tilde{i}, \tilde{j}, \tilde{k})}^{mem} < C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{mem}

(27)

0 \leq d_{(\tilde{i}, \tilde{j}, \tilde{k})}^{bw} < C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{bw}

(28)

T_{jo int}^{(\tilde{i}, \tilde{j}, \tilde{k})} \leq T_{th}

(29)

i_{\tilde{j}} = i_{\tilde{j} - 1} - i_{\tilde{j} - 1}^{finish} + i_{\tilde{j} - 1}^{arrive}

(30)

i_{total} = \sum_{\tilde{j} = 1}^{j} i_{\tilde{j} - 1}^{arrive}

(31)

The purpose of the formula (23) is to find the minimum energy cost, where $E_{map - level}^{total}$ and $E_{clo - level}^{total}$ can be calculated by equations (12) and (14), respectively. The purpose of equation (24) is to find the maximum reliability. The constraint (equation (25)) ensures that the numbers are nature number. Formulas (26), (27), and (28) indicate that the resources such as CPU, memory, and bandwidth should be less than the actual available resources, where $d_{(\tilde{i}, \tilde{j}, \tilde{k})}^{cpu}$ is the consumption of CPU, $d_{(\tilde{i}, \tilde{j}, \tilde{k})}^{mem}$ is the consumption of memory, and $d_{(\tilde{i}, \tilde{j}, \tilde{k})}^{bw}$ is the consumption of bandwidth. $C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{cpu}$ , $C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{mem}$ , and $C_{(\tilde{i}, \tilde{j}, \tilde{k}, \tilde{m})}^{bw}$ are the available resources of system that are calculated by equations (15), (16), and (17), respectively. Constraint (equation (29)) ensures that task execution time should be not more than the time threshold. In sum, the function of formulas (26), (27), (28), and (29) ensure system performance and user satisfaction. Equations (30) and (31) are used to explain the change of the number of users, where $i_{\tilde{j}}$ is the number of users at time $\tilde{j}$ , $i_{\tilde{j} - 1}$ is the number of users at time $\tilde{j} - 1$ , $i_{\tilde{j} - 1}^{finish}$ is the number of users that was completed by system at time $\tilde{j} - 1$ , and $i_{\tilde{j} - 1}^{arrive}$ is the number of new users arriving in system at time $\tilde{j} - 1$ . Meanwhile, if the system adds the user to the user set $U$ when the system receives a new user request, the total number of user $i_{total}$ is equal to the size of $U$ .

Optimization algorithm based on GA

At present, the heuristic method is a common method to search global optimization solution, such as Simulated Annealing (SA),⁴¹ Ant Colony Optimization (ACO),⁴² Tabu Search (TS),⁴³ Particle Swarm Optimization (PSO),⁴⁴ and GA.⁴⁵ The SA algorithm is easy to fall into the local search, so it is not appropriate to find the optimal solution. The TS algorithm provides memory capacity to record the previous action, which can avoid the emergence of the local optimal solution, but it may cause a lot of memory waste. The ACO algorithm is not suitable for recursive prediction, because it has higher time complexity. The GA can solve the complex optimization problem by using crossover, mutation, and selection operation. More importantly, the GA is more efficient and faster than brute force search algorithm. Therefore, genetic algorithm based on time series matching method (GATM) for optimal energy consumption is presented in this section. The algorithm integrates time-series matching method and GA to achieve energy saving on the premise of guaranteeing the quality of service.

Algorithm. GATM.
${CSM}_{\tilde{i}}$ : category of service mode of $\tilde{i} th$ user; $Q_{\tilde{i}}$ : the historical trajectory of $\tilde{i} th$ user; $q_{\tilde{i}}^{\tilde{j}, \tilde{k}}$ : $\tilde{i} th$ user request; $q_{\overset{⌣}{i}}^{\tilde{j}, \tilde{k}}$ : the latest history request that is consistent with $q_{\tilde{i}}^{\tilde{j}, \tilde{k}}$ ; $D_{mp}$ : mapping-level service mode database; $D_{cl}$ : cloud-level service mode database; $P_{c}$ : crossover probability; $P_{m}$ : mutation probability;
1:	Begin
2:	Input: $q_{\tilde{i}}^{\tilde{j}, \tilde{k}}$ , $D_{map}$ , $D_{clo}$ , $P_{c}$ , $P_{m}$
3:	For $q_{\tilde{i}}^{\tilde{j}, \tilde{k}}$
4:	Getting ${CSM}_{\tilde{i}}$
5:	If ${CSM}_{\tilde{i}} = 0$
6:	$q_{\tilde{i}}^{\tilde{j}, \tilde{k}}$ will be executed by the mapping-level service mode
7:	Adding $q_{\tilde{i}}^{\tilde{j}, \tilde{k}}$ to $D_{mp}$
8:	Else if ${CSM}_{i} = 1$
9:	$q_{\tilde{i}}^{\tilde{j}, \tilde{k}}$ will be executed by the cloud-level service mode
10:	Adding $q_{\tilde{i}}^{\tilde{j}, \tilde{k}}$ to $D_{cl}$
11:	Else
12:	Using time-series matching algorithm to find $q_{\overset{⌣}{i}}^{\tilde{j}, \tilde{k}}$ in the historical database
13:	If $q_{\overset{⌣}{i}}^{\tilde{j}, \tilde{k}}$ is existed in the historical database
14:	Getting $d_{(\overset{⌣}{i}, \overset{⌣}{j}, \overset{⌣}{k})}^{c p u}$ , $d_{(\overset{⌣}{i}, \overset{⌣}{j}, \overset{⌣}{k})}^{m e m}$ , $d_{(\overset{⌣}{i}, \overset{⌣}{j}, \overset{⌣}{k})}^{b w}$ , $T_{th}$
15:	Getting $C_{(\tilde{i}, \tilde{j}, \tilde{m})}^{cpu}$ , $C_{(\tilde{i}, \tilde{j}, \tilde{m})}^{mem}$ , $C_{(\tilde{i}, \tilde{j}, \tilde{m})}^{bw}$ with the help of control monitor
16:	If $d_{(\overset{⌣}{i}, \overset{⌣}{j}, \overset{⌣}{k})}^{c p u} < C_{(\tilde{i}, \tilde{j}, \tilde{m})}^{c p u}$ , $d_{(\overset{⌣}{i}, \overset{⌣}{j}, \overset{⌣}{k})}^{m e m} < C_{(\tilde{i}, \tilde{j}, \tilde{m})}^{m e m}$ , $d_{(\overset{⌣}{i}, \overset{⌣}{j}, \overset{⌣}{k})}^{b w} < C_{(\tilde{i}, \tilde{j}, \tilde{m})}^{b w}$ , and $T_{map - level} < T_{th}$
17:	Adding $q_{\tilde{i}}^{\tilde{j}, \tilde{k}}$ to $D_{mp}$
18:	Else
19:	Adding $q_{\tilde{i}}^{\tilde{j}, \tilde{k}}$ to $D_{cl}$
20:	Else
21:	Adding $q_{\tilde{i}}^{\tilde{j}, \tilde{k}}$ to $D_{cl}$
22:	Start encoding for task
23:	Initialization solution
24:	Evaluating solution according to $F (E)$ and $F (I)$
25:	If termination condition is satisfied
26:	Performing (33)
27:	Else
28:	Performing the selection strategy based on roulette
29:	Performing the crossover operators according to $P_{c}$
30:	Performing the mutation operators according to $P_{m}$
31:	Generating new solution
32:	Repeating (24)–(31)
33:	Print optimal solution
34:	End

The core of this algorithm is using historical trajectory to provide guidance for future service mode selection. Steps (3)–(21) are to perform a preliminary service mode selection. Steps (22)–(33) are to find the optimal solution by GA. More specifically, steps (5)–(10) are to execute service mode selection according to users’ personalized needs. Steps (11)–(21) are to execute service mode selection according to system capability and historical record. After completing the steps described above, we can obtain mapping-level service mode database and cloud-level service mode database. Step (22) is to encode candidate solutions, which are represented in binary as strings of 0s and 1s. Step (23) is to generate an initial population randomly. Step (24) is to evaluate the fitness of every solution. Steps (28)–(31) select the best solutions and execute crossover, mutation strategy to form a new solution. Step (32) is to repeat the above operation until the termination condition is satisfied. Step (33) is to print the optimal solution. If the process is terminated due to a maximum number of generations, a satisfactory solution may or may not have been reached.

Simulation and analysis

In this section, we evaluate the performance of the task joint execution strategy from three aspects: the matching accuracy ratio of time-series matching method, mobile terminal energy consumption, and task success ratio. The experimental parameters and experimental results are given as follows.

Simulation environment and parameters

To evaluate the performance of the GATM method, we choose three methods for comparison. They are GA method, CE method, and local execution (LE) method. The GATM method and GA method are task joint execution strategy. The CE method is a cloud execution strategy. The LE method is a local execution strategy. For CE method, the service is provided by the cloud-level service mode. For LE method, the service is provided by the mapping-level service mode. The GATM, GA, CE, and LE methods are implemented in Python language. The performances are measured in PC with Intel Core i3-2130 3.40 GHz processor and 4.00 GB memory. Simulation parameters are $α = β = γ = 1 / 3$ , $R_{mp} = 1$ , $R_{cl} = 1.5$ , $P_{mp}^{power} = 2$ , $P_{cl}^{power} = 1$ , $P_{off}^{power} = 1$ , $R_{off} \in [0, 1]$ , $P_{c} = 0.9$ , and $P_{m} = 0.2$ , and others are random generated.

Numerical results

In this section, we evaluate the performance of the time-series matching method, and the results are shown in Figures 4 and 5. In addition, the results of GATM, GA, CE, and LE methods are shown in Figures 6 and 7.

Figure 4.

The accuracy ratio versus sample size.

Figure 5.

The accuracy ratio versus the shortest match length.

Figure 6.

Terminal energy consumption, the number of Cloudlet, and task success ratio of different methods.

Figure 7.

Terminal energy consumption, latest finishing time, and task success ratio of different methods.

It can be seen from Figure 4, with the increasing of sample size, the accuracy ratios of the time-series matching method and the Markov method are increasing gradually. In the initial stage, the accuracy ratios of the time-series matching method and the Markov method increase rapidly. After the number of sample size reaches a specific value, the accuracy will keep a stable value approximately. But the time-series matching method has higher accuracy than the Markov method. For example, the maximum accuracy ratio of time-series matching method is 85.181% and that of Markov method is 75.865%. The minimum accuracy ratios of the time-series matching method and the Markov method are 82.562% and 70.463%, respectively. More importantly, the maximum accuracy ratio and minimal accuracy ratio of time-series matching method increase by 9.316% and 12.099%, respectively. The results show that the time-series matching method has better performance. The main reason for this phenomenon is that the Markov method equally treats every task, and Markov process dissevers the internal correlation between user requests.

Figure 5 shows the relationship between accuracy ratio and the match length. With the increasing of match length, the accuracy ratios of the short interest model and the long interest model are increasing gradually. After match length reaches a specific value, the accuracy will keep a stable value approximately. The accuracy ratio of the short interest model increases to 95.253% from 80.049% and that of the long interest model increases to 87.136% from 63.043%. This proves that the short interest model is effective in mobile micro-learning. The main reason is that the short interest model has a more powerful influence on the next requests.

The terminal energy consumption and the task success ratio of different methods are shown in Figures 6 and 7. In Figure 6, the constraint condition is the number of Cloudlet. In Figure 7, the constraint condition is the latest finishing time.

The relationships among terminal energy consumption, the number of Cloudlet server, and task success ratio are shown in Figure 6. For the same number of Cloudlet, the simulation results show that the GATM method is effective in energy saving and reliability. The realistic evidence is that the terminal energy consumption of GATM is lower and its fluctuation is less. Conversely, terminal energy consumption that calculated by GA method and CE method are not only high but also volatile. For LE method, its energy consumption has been maintaining at a relatively stable level. Based on the above analysis, we think that the GATM method is more advantageous in terms of energy saving. At the same time, the task success ratio of GATM method is significantly higher than that of the GA, CE, and LE methods. The GATM method maintains a high task success ratio that is between 75% and 90%. Meanwhile, the task success ratio of the GA method is lower than 80%. The task success ratio of the CE method is less than 20%. For LE method, it has the high task success ratio, but its energy consumption is also high. The above analysis proved the GATM method is energy saving and reliability.

To further evaluate the performance of our method, the average energy consumption (AEC) and average task success rate (ATSR) are used to measure the energy saving and reliability. The experimental results are shown in Table 1.

Table 1.

AEC and ATSR under the constraint of the number of Cloudlet.

Algorithm	AEC (mj)	ATSR (%)
GATM	30.4158	0.7840
GA	68.5815	0.7082
CE	503.7787	0.0635
LE	169	1

AEC: average energy consumption; ATSR: average task success rate; GATM: genetic algorithm based on time series matching method; GA: genetic algorithm; CE: cloud execution; LE: local execution.

From Table 1, we can see that the AEC and ATSR of the GATM method are obviously higher than other methods. The AEC and ATSR of GATM method are 30.4158 mj and 78.40%, respectively, which is higher than that obtained by the GA, CE, and LE methods. Therefore, GATM method achieves the goal of energy efficiency. Meanwhile, the variance of the GATM, GA, CE, and LE methods is 320.1352, 7780.8759, 9503.7129, and 0, respectively. This proves that the GATM method is stable in energy saving. The task success ratio of the GATM method is 0.7840, which is 0.0758 higher than that of the GA method and 0.7205 than the LE method. At the same time, the variance of those methods is 0.00101376, 0.016899178, 0.00150530, and 0, respectively. Given the above, GATM method has better performance in reducing energy consumption and ensuring service reliability than other methods do.

The relationships among terminal energy consumption, latest finishing time, and task success ratio are shown in Figure 7. When users have same latest finishing time, GATM method achieves the trade-off between energy consumption and task success ratio. The terminal energy consumption of GATM method is 0 to 50 mj, and its task success ratio is 60% to 80%. The task success ratio of the GA method is relatively high, but its terminal energy consumption is complex and volatile. Although CE method achieves energy saving, the task success ratio is relatively low. The LE method has the highest task completion rate, but it also has the highest energy consumption. Therefore, experimental results prove that GATM method is efficient in energy saving and reliability.

To show the green and reliability of the GATM, GA, CE, and LE methods, AEC and ATSR of the GATM, GA, CE, and LE methods with the limitation of the latest finish time are shown in Table 2.

Table 2.

AEC and ATSR under the constraint of the latest finish time.

Algorithm	AEC (mj)	ATSR (%)
GATM	30.1188	0.7515
GA	291.8555	0.6502
CE	182.8367	0.0288
LE	338	1

AEC: average energy consumption; ATSR: average task success rate; GATM: genetic algorithm based on time series matching method; GA: genetic algorithm; CE: cloud execution; LE: local execution.

From Table 2, we can see that the AEC of GATM method is 30.1188, that of GA method is 291.8555, that of CE method is 182.8367, and that of LE method is 338 mj, respectively. Meanwhile, the variance of GATM is 13.9422, that of GA is 7780.8759, and that of CE is 182.8367. Thus, the AEC indicates that GATM method is energy saving. More importantly, the ATSR of GATM, GA, CE, and LE methods is 0.7515, 0.6502, 0.0635, and 1, respectively, which prove the GATM method is reliable. Therefore, GATM method is energy saving and dependability.

The experimental results of Figures 6 and 7 show that GATM method can achieve the goal of energy efficiency and ensure the reliable execution of the task. It is largely thanks to three processes that are time-series matching, service mode selection, and GA optimization. The accuracy ratio of time-series matching method can provide guidance for the initial service mode selection. The service mode selection module provides two service modes, which can achieve seamless switching by offloading operation. In GA module, the task is actually performed with the limitations of the network environment and resource status. Therefore, even in the case of a matching error, the optimal task execution strategy can be found through the service mode selection module or the GA optimization module.

Conclusion and future works

In this article, a task joint execution strategy is proposed to reduce the energy consumption of mobile terminal in mobile micro-learning. First, we mine the latest trajectory of the service from the historical database according to the sequential and relative stability of mobile micro-learning. Second, the mapping-level service mode and cloud-level service mode are provided to achieve seamless switching based on resource and users’ personalized need. Finally, to achieve energy saving target, we adopt the GATM method to find a suitable service mode. The simulation results show that the task joint execution strategy can reduce the mobile terminal energy consumption and ensure the dependability of mobile micro-learning within the threshold time. In the future work, we will focus on the optimal match length of the time-series matching method, the precise location of the users, selfishness of Cloudlet, additional energy consumption, time cost caused by other operation, and so on.

Footnotes

Handling Editor: Aitor Almeida

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work is partially supported by the National Natural Science Foundation of China (NSFC) under grant nos 61602155, 61370221, and U1604155; in part by Henan Science and Technology Innovation Project under grant nos 164200510007 and 174100510010; in part by the Program for Science & Technology Innovation Talents in the University of Henan Province under grant no. 16HASTIT035; in part by the Key Scientific and Technological Projects Henan Province under grant no. 172107000005; and in part by the support program for young backbone teachers in Henan Province under grant no. 2015GGJS-047.

References

Sou

Peng

. Performance modeling for multipath mobile data offloading in cellular/Wi-Fi networks. IEEE T Commun 2017; 65(9): 3863–3875.

Sun

Hou

Zhou

et al . Equilibriums in the mobile-virtual-network-operator-oriented data offloading. IEEE T Veh Technol 2018; 67(2): 1622–1634.

Park

Song

. Game-theoretic scalable offloading for video streaming services over LTE and WiFi networks. IEEE T Mobile Comput 2017; 17(5): 1090–1104.

Eriksson

Dan

Fodor

. Coordinating distributed algorithms for feature extraction offloading in multi-camera visual sensor networks. IEEE T Circ Syst Vid. Epub ahead of print 25 August 2017. DOI: 10.1109/TCSVT.2017.2745102.

Kamaruzaman

Zainol

. Behavior response among secondary school students development towards mobile learning application. In: IEEE colloquium on humanities, science and engineering, Kota Kinabalu, Malaysia, 3–4 December 2012, pp.589–592. New York: IEEE.

Kolb

. Learning styles and learning spaces: enhancing experiential learning in higher education. Acad Manag Learn Edu 2005; 4(2): 193–212.

Tam

Lam

Fung

T. Toward a complete e-learning system framework for semantic analysis, concept clustering and learning path optimization. In: 12th IEEE international conference on advanced learning technologies, Rome, 4–6 July 2012, pp.592–596. New York: IEEE.

Khalil

Ebner

. MOOCs completion rates and possible methods to improve retention—a literature review. In: World conference on education multimedia, hypermedia and telecommunications, Tampere, 23 June 2014, pp.1236–1244. Waynesville, NC: AACE.

Gungor

Sahin

Kocak

et al . Smart grid technologies: communication technologies and standards. IEEE T Ind Inform 2011; 7(4): 529–539.

10.

Eichi

Zeng

et al . A survey on the electrification of transportation in a smart grid environment. IEEE T Ind Inform 2012; 8(1): 1–10.

11.

Chourabi

Nam

Walker

et al . Understanding smart cities: an integrative framework. In: 45th Hawaii international conference on system science, Maui, HI, 4–7 January 2012, pp.2289–2297. New York: IEEE.

12.

Kotevska

Kusne

Samarov

et al . Dynamic network model for smart city data-loss resilience case study: city-to-city network for crime analytics. IEEE Access 2017; 5: 20524–20535.

13.

Giovannella

. Territorial smartness and the relevance of the learning ecosystems. In: 1st IEEE international smart cities conference, Guadalajara, Mexico, 25–28 October 2015, pp.1–5. New York: IEEE.

14.

Markovic

Sofronijevic

. Building a gamified system for capturing MOOC related data: smart city learning community as its most precious source of intangible cultural heritage. In: International conference on culture and computing, Kyoto, Japan, 17–19 October 2015, pp.175–182. New York: IEEE.

15.

Salerno

. Knowledge lifecycle and smart cities learning position paper for DUBAI 2020: smart city learning. In: International conference on web and open access to learning, Dubai, United Arab Emirates, 25–27 November 2014, pp.1–4. New York: IEEE.

16.

Liu

. English teaching model research based on network MOOC. In: International conference on smart city and systems engineering, Hunan, China, 25–26 November 2016, pp.183–186. New York: IEEE.

17.

Malek

Laroussi

Ghezala

. A design framework for smart city learning scenarios. In: 9th international conference on intelligent environments, Athens, 16–17 July 2013, pp.9–15. New York: IEEE.

18.

Elhoseny

Abdelrazek

et al . Ubiquitous smart learning system for smart cities. In: 8th international conference on intelligent computing and information systems, Cairo, Egypt, 5–7 December 2017, pp.329–334. New York: IEEE.

19.

Hammad

Ludlow

. Towards a smart learning environment for smart city governance. In: 9th IEEE/ACM international conference on utility and cloud computing, Shanghai, China, 6–9 December 2016, pp.185–190. New York: IEEE.

20.

Kadar

. Smart learning environment for the development of smart city applications. In: 8th IEEE international conference on intelligent systems, Sofia, 4–6 September 2016, pp.59–64. New York: IEEE.

21.

Yin

Zhang

Rong

et al . Mobile learning for professional situations in smart city. In: IET international conference on smart and sustainable city, Shanghai, China, 19–20 August 2013, pp.401–406. New York: IEEE.

22.

Souza

MIF

Amaral

SFD

. Educational microcontent for mobile learning virtual environments. Creat Educ 2014; 5(9): 672–681.

23.

Sun

Cui

Chen

et al . MLaaS: a cloud system for mobile micro learning in MOOC. In: IEEE international conference on mobile services, New York, 27 June–2 July 2015, pp.120–127. New York: IEEE.

24.

Kim

Mims

Holmes

. An introduction to current trends and benefits of mobile wireless technology use in higher education. AACE J 2006; 14(1): 77–100.

25.

Chen

Zhang

Sun

et al . Personalized micro-learning support based on process mining. In: 7th international conference on information technology in medicine and education, Huangshan, China, 13–15 November 2015, pp.511–515. New York: IEEE.

26.

Fekete

Caorba

Forstner

et al . Energy-efficient computation offloading model for mobile phone environment. In: 1st IEEE international conference on cloud networking, Paris, 28–30 November 2012, pp.95–99. New York: IEEE.

27.

Shu

Liu

Jin

et al . eTime: energy-efficient transmission between cloud and mobile devices. In: IEEE international conference on computer communications, Turin, 14–19 April 2013, pp.195–199. New York: IEEE.

28.

Haridas

Kailasam

Dharanipragada

. Cloudy Knapsack algorithm for offloading tasks from large scale distributed applications. IEEE T Cloud Comput. Epub ahead of print 4 July 2017. DOI: 10.1109/TCC.2017.2713776.

29.

Chatzopoulos

Ahmadi

Kosta

et al . FlopCoin: a cryptocurrency for computation offloading. IEEE T Mobile Comput 2017; 17: 1062–1075. DOI: 10.1109/TMC.2017.2748133.

30.

Ahmed

. A survey on mobile edge computing. In: 10th international conference on intelligent systems and control, Coimbatore, India, 7–8 January 2016, pp.1–8. New York: IEEE.

31.

Intharawijitr

Iida

Koga

et al . Practical enhancement and evaluation of a low-latency network model using mobile edge computing. In: 41st IEEE annual computer software and applications conference, Turin, 4–8 July 2017, pp.576–574. New York: IEEE.

32.

Wei

Wang

Zhou

et al . MVR: an architecture for computation offloading in mobile edge computing. In: IEEE international conference on edge computing, Honolulu, HI, 25–30 June 2017, pp.232–235. New York: IEEE.

33.

Chen

Ren

. Online learning for offloading and autoscaling in energy harvesting mobile edge computing. IEEE T Cogn Commun Netw 2017; 3(3): 361–373.

34.

Wang

Liang

et al . Joint computation offloading and interference management in wireless cellular networks with mobile edge computing. IEEE T Veh Technol 2017; 66(8): 7432–7445.

35.

Cao

Cai

. Distributed multiuser computation offloading for Cloudlet-based mobile cloud computing: a game-theoretic machine learning approach. IEEE T Veh Technol 2018; 67(1): 752–746.

36.

Abd

Al- Haddad

SAR

Hashim

et al. Energy-aware fault tolerant task offloading of mobile cloud computing. In: 5th IEEE international conference on mobile cloud computing, services, and engineering, San Francisco, CA, 6–8 April 2017, pp.161–164. New York: IEEE.

37.

Lippi

Bertini

Frasconi

. Short-term traffic flow forecasting: an experimental comparison of time-series analysis and supervised learning. IEEE T Intell Transp 2013; 14(2): 871–882.

38.

Mistry

Bouguettaya

et al . Long-term QoS-aware cloud service composition using multivariate time series analysis. IEEE T Serv Comput 2016; 9(3): 382–393.

39.

Wang

Liu

Zhang

. Access control based resource allocation in cloud computing environment. Int J Netw Secur 2017; 19(2): 236–243.

40.

Ding

Maharjan

et al . Decentralized and optimal resource cooperation in geo-distributed mobile cloud computing. IEEE T Emerg Top Comput 2015; 6(1): 72–84.

41.

Kirkpatrick

Gelatt

Vecchi

. Optimization by simulated annealing. Science 1983; 220(4598): 671–680.

42.

Marco

Montes

DOMA

Sabrina

et al . Ant colony optimization. IEEE Comput Intell M 2004; 1(4): 28–39.

43.

Glover

. Tabu search: a tutorial. Interfaces 1990; 20(4): 74–94.

44.

Kennedy

Eberhart

. Particle swarm optimization. In: IEEE international conference on neural networks, Perth, WA, Australia, 27 November–1 December 1995, pp.1942–1948, vol. 4. New York: IEEE.

45.

Shahidehpour

Bahramirad

et al . Optimizing traffic signal settings in smart cities. IEEE T Smart Grid 2017; 8(5): 2382–2393.

Green city: An efficient task joint execution strategy for mobile micro-learning

Abstract

Keywords

Introduction

Related work

Task joint execution strategy modeling

Mobile micro-learning system architecture

Task joint execution strategy model

Problem formulations

Transforming mobile micro-learning process into time-series model

Time-series matching method

Utility function

Optimization algorithm based on GA

Simulation and analysis

Simulation environment and parameters

Numerical results

Conclusion and future works

Footnotes

Declaration of conflicting interests

Funding

References