Machine Learning-Based Radio Access Technology Selection in the Internet of Moving Things

Ramon Sanchez-Iborra,Luis Bernal-Escobedo,Jose Santa

1 Dept.Engineering and Applied Techniques,University Centre of Defence at the Spanish Air Force Academy,30729 San Javier,Spain

2 Dept.Information and Communication Engineering,University of Murcia,30100 Murcia,Spain

3 Dept.of Electronics,Computing Technology and Projects,Technical University of Cartagena,30202 Cartagena,Spain

Abstract: The Internet of Moving Things (IoMT)takes a step further with respect to traditional static IoT deployments.In this line,the integration of new eco-friendly mobility devices such as scooters or bicycles within the Cooperative-Intelligent Transportation Systems(C-ITS)and smart city ecosystems is crucial to provide novel services.To this end,a range of communication technologies is available,such as cellular,vehicular WiFi or Low-Power Wide-Area Network(LPWAN);however,none of them can fully cover energy consumption and Quality of Service (QoS) requirements.Thus,we propose a Decision Support System(DSS),based on supervised Machine Learning(ML) classification,for selecting the most adequate transmission interface to send a certain message in a multi-Radio Access Technology(RAT)set up.Different ML algorithms have been explored taking into account computing and energy constraints of IoMT enddevices and traffic type.Besides,a real implementation of a decision tree-based DSS for micro-controller units is presented and evaluated.The attained results demonstrate the validity of the proposal,saving energy in communication tasks as well as satisfying QoS requirements of certain urgent messages.The footprint of the real implementation on an Arduino Uno is 444 bytes and it can be executed in around 50µs.

Keywords:internet of moving things; multi-RAT;CITS;classification;personal mobility

I.INTRODUCTION

With the arrival of the Internet of Moving Things(IoMT) [1]a massively connected infrastructure as well as uncountable sources of data become available.This network permits to connect every world’s corner and consists of sensing/actuator devices that move with users,e.g.,wearables and e-health gadgets,or move attached to monitoring units such as road furniture or public vehicles,and others that move on their own accord,e.g.,robots and autonomous vehicles.In this line,mobility trends in metropolitan areas are currently evolving due to the emergence of personal transportation means,e.g.,electric mopeds,shared bicycles,different type of scooters,etc.These vehicles are being adopted by citizens because of their eco-friendly characteristics and their mobility advantages in urban scenarios,which permit to avoid traffic jams at the time that promote healthy lifestyles.Besides,their inclusion within hyper-connected spaces such as those under the umbrella of smart cities or the Cooperative-Intelligent Transportation Systems (C-ITS) paradigm is becoming a reality thanks to the connectivity capabilities of On-Board Units (OBUs).These elements can be equipped with different communication technologies,which is known as multi-Radio Access Technology (RAT),aiming at gaining access to a plethora of cloud services such as traffic information,vehicle tracking and monitoring,route planning,etc.(Figure 1).Examples of typical RATs for vehicular scenarios are vehicular WiFi,cellular networks,or Low Power Wide Area Network (LPWAN),among others[2].Given the power constraints of personal vehicles,the RAT selection for transmitting a given message should be carefully addressed to optimize power consumption,as some communication technologies are more energetically efficient than others.Additionally,certain messages present Quality of Service(QoS)requirements in terms of latency or link reliability that some RATs are unable to provide.Therefore,it is clear that effective Decision Support Systems(DSSs)are needed to be embedded into OBUs to select the optimal RAT for every message to be transmitted[3].

Figure 1.Multi-RAT OBU and services for personal mobility devices.

DSSs have notably improved their performance thanks to the application of a variety of Machine Learning (ML) techniques,e.g.,Artificial Neural Networks (ANN),Support Vector Machines (SVM),Bayes Learning (BL),Random Forest (RF),etc.[4].These algorithms are adopted to make the most appropriate decision according to a series of restrictions and optimization objectives in a wide range of use cases and scenarios.For that reason,this work bets on the use of ML-based DSSs as a proper solution for selecting the most adequate RAT at any given time.In fact,the recent appearance of the TinyML paradigm [5],which proposes to adapt ML algorithms to constrained processing platforms such as Micro Controller Units(MCUs),opens the door for unimaginable developments in scarce-resource devices.

To the best of authors’ knowledge,there is not any prior work addressing the challenge of intelligent RAT selection in the field of IoMT and,concretely,in the segment of personal mobility.In this paper we tackle this issue as a classification problem in which we consider both the status of the“thing”and its incorporated RATs,as well as the characteristics of the message to be sent.We have trained a number of classification algorithms and,by means of computer simulation,we have evaluated them in terms of achieved accuracy,OBU’s energy efficiency and communication QoS.Finally,the most accurate and efficient method has been implemented in an Arduino Uno development board and its performance has been studied.The main contributions of this work are the following:

• A multi-RAT OBU architecture for eco-friendly IoMT devices.

• A multi-RAT selection solution based on supervised ML multi-class classification,addressing OBU constraints and traffic QoS.

• A real DSS implementation for Arduino based on decision tree algorithm.

The remaining of the paper is organized as follows: Section II reviews related works in the literature and identifies the advances beyond the state of the art.Section III presents the general architecture of the multi-RAT solution for personal mobility devices.Section IV details the ML-based multi-RAT decision scheme and its implementation.Section V addresses the evaluation of the proposal.Finally,Section VI concludes the paper and proposes future research directions.

II.RELATED WORK

As mentioned above,to the best of our knowledge,there is no prior work addressing the multi-RAT selection issue in the field of smart and sustainable transportation.However,some related efforts have been done in the area of IoT.Work in [3]provided a wide overview of multi-RAT management in massive Machine-Type Communications (mMTC) environments.Authors exclusively focused on the range of LPWAN-based technologies,concretely on Lo-RaWAN and Narrow Band-Internet of Things (NBIoT).A prototype incorporating one transceiver for each of the considered RATs was implemented and the energy consumption of the module was characterized.However,in this work there was not presented any RAT selection procedure,but the results only reported the power consumption of each RAT depending on message lengths.From a global system perspective,authors of[6]studied the scenario of multiple devices simultaneously communicating through several independent channels.A total power minimization problem was formulated according to the Shannon capacity theorem with power and Signal-to-Interference-plus-Noise Ratio (SINR) constraints.The paper results,obtained by computer simulation,showed a clear improvement in the overall power efficiency of the system.

Work in [7]presented a reinforcement learningbased approach for deciding the RAT to be employed by a multi-RAT IoT-device.Authors considered 5G and LoRa as the available RATs,and proposed a reinforcement algorithm’s reward for maximizing the throughput of transmissions.Simulation results revealed a good performance of the mechanism,although it was not implemented over real constrained hardware.In [8],another reinforcement learningbased algorithm was presented for solving the problem of intelligent base-station selection.Specifically,authors focused on mMTC devices in order to enablethem to cooperate for minimizing network congestion.Additional works have also considered the use of ML mechanisms for deciding the best RAT to be employed but they were oriented to pure 5G scenarios [9,10],hence ignoring the constraints of IoT-based systems.

Besides the multi-RAT selection problem,the integration of ML-based intelligence within IoMT enddevices is gaining great momentum thanks to the arrival of the TinyML paradigm[5].It proposes to adapt ML models generated in non-constrained architectures to make then runnable by constrained end-devices.Furthermore,the alliance of TinyML with edge computing paves the way for the development of novel services devoted to IoT deployments,e.g.,radio access management,data caching,task offloading,or digital twins,among many others [11].Given the recent creation of the TinyML ecosystem,few works have exploited its promising possibilities in IoMT scenarios.Work in [12]presented a Convolutional Neural Network (CNN)-based mechanism for improving the performance of mini-vehicles’autonomous driving by means of on-device accurate image classification.In turn,work in[13]presented a TinyML-powered alcohol sensor embedded in a wearable device.With this approach,authors aimed to reduce dependency on the network connectivity,which is generally bandwidth and power consuming,also resulting in higher latencies.Besides,the TinyML approach permits to preserve user/data confidentiality since data is processed at the device itself.

Different from previous work,we focus on the specific problem of multi-RAT selection for highly constrained IoMT devices,concretely for the case of personal mobility vehicles.This is a relevant issue given the high mobility of these elements,which may cause intermittent coverage regions for specific RATs that should be complemented with the use of others.To this end,we evaluate the accuracy of a series of MLbased classification algorithms and,finally,a real implementation over a well-known MCU board is carried out.The vision and developments of this work may be impactful across several areas of study aiming at fostering the design,implementation,and deployment of future multi-RAT constrained solutions.

Then they applied21 for advice to one of the wise men of the world,and he sent them a message to say that there was one remedy whichwould relieve and cure him, and that it was a plant of heavenly originwhich grew in the forest in the king s own dominions22. The messengerdescribed the flower so that is appearance could not be mistaken.Then said the swineherd, I am afraid I carried this plant awayfrom the forest in my bundle, and it has been burnt to ashes long ago.But I did not know any better.

III.GENERAL ARCHITECTURE

The proposed multi-RAT OBU architecture for personal mobility devices is presented in Figure 2.As observed,it consists of two principal elements: the main board and the communication concentrator.Regarding the main board,it is equipped with a MCU and memory,which can be integrated in a System on Chip(SoC) fashion.The processing power of this element is certainly limited,as energy efficiency is prevalent in this type of solutions.Inside the MCU,the DSS algorithm code is run in order to manage the RAT selection switch placed in the communication concentrator,as further explained below.The main board also mounts a range of sensors not limited to those showed in the diagram,in order to collect data from the vehicle and the environment.A GPS is also useful in this type of devices for tracking or monitoring purposes.Finally,this board presents an external display to output status messages and an external battery that cannot be of large capacity for avoiding an excessive payload in terms of size and weight.

Figure 2.Overall architecture for the DSS-assisted multi-RAT OBU.

Figure 3.Labeled samples from the generated data set(5000 samples).

The communication concentrator,which provides the main board with multi-RAT capabilities,is equipped with three of the most employed communication technologies in urban vehicular scenarios,as mentioned above: LPWAN,e.g.,LoRaWAN or NBIoT;vehicular WiFi,including IEEE 802.11 OCB,formerly known as 802.11p; and cellular networks,e.g.,4G/5G.These three technologies offer different capabilities that properly combined improve application needs and user experience.4G/5G can imply relatively high power consumption,but its almost ubiquitous access is useful at particular areas lacking LPWAN and WiFi coverage.Therefore,the respective transceivers are integrated within this module,which also presents a RAT selector switch that is commanded from the main board.Having this architecture in mind,it is considered that the MCU runs a piece of software which consists of an infinite loop that(i)collects sensor data and captures unexpected events,(ii)manipulates these data,(iii) decides when and how to report this information,and,finally,(iv) sends the message.This is a common work flow in IoT monitoring systems[14].For the transmission decision process,the DSS evaluates the status of the OBU as well as the QoS characteristics of the message to be sent and decides the most proper interface to be employed.This process is explained in detail in Section IV.

Once the most adequate RAT has been selected,the message is sent over the air to the corresponding in-frastructure’s point of contact,i.e.,gateway,Road Side Unit (RSU),or base station,depending on the transmission technology.Then,these elements forward the data through their respective backhaul networks to the cloud servers in the Internet.Although LPWAN technologies such as LoRaWAN could involve the direct transmission of messages above a link-level layer,the middleware developed in [15]allows the homogeneous treatment of these links as regular IPv6 interfaces.Finally,different services may be accessed by users through their personal smart devices.Hence,different applications such as vehicle tracking,route planning,or emergency event management can be efficiently developed.These services present different QoS requirements,specially in terms of latency or communication reliability,which has a direct impact on the decision made by the OBU advised by the DSS.

IV.ML-BASED MULTI-RAT SELECTION

ML techniques are being constantly refined and nowadays many of them can be deployed and executed in highly constrained processing hardware like MCUs[16].These devices present a highly reduced power consumption and they are provided with a series of digital and analog ports to interact with a range of sensors and actuators.Therefore,these units are good candidates to be adopted as the processing core of personal mobility vehicle’s OBUs.Besides,the adaptation of well-known ML libraries,e.g.,TensorFlow Lite [17]or the development of specific libraries for these devices,e.g.,Microsoft’s Embedded Learning Library [18],is paving the way for the integration of Artificial Intelligence(AI)and the development of DSS blocks within MCUs.

4.1 Problem Statement and DSS Definition

Let a personal vehicle’s OBU monitor certain parameters of interest that should be delivered to their corresponding application servers in the cloud.The OBU is provided with a number of RATs that can be employed for transmitting retrieved data from the embedded sensors.When some data is collected from a sensor or an unexpected event is captured,the OBU should decide whether to report it or not.In the case of deciding to send the message,it has to choose the most adequate RAT to perform this action.Thereby,the set of all possible actions(A)is{a0,a1,...,an},wherea0represents the action of dropping the data and{a1,...,an}indicates sending the data using one of thenavailable RATs.This decision is made by the DSS module,by evaluating certain features which define the status of both the OBU and the data to be sent.

Regarding the OBU’s status,three different aspects are considered.Firstly,the battery level(B),which is of prominent importance as mentioned above.Hence,B reports the percentage of remaining energy.In this regard,each actionairequires certain amount of energy (E) in order to be performed,expressed by{e0,e1,...,en}.The coverage level (C) of each RAT is also considered as{c1,c2,...,cn},withi≠ 0,aiming at addressing transmission reliability.Besides,the data-rate (R) of each RAT has been also taken into account as{r1,r2,...,rn}withi≠ 0.Finally,some RATs make use of license-free spectrum,e.g.,Industrial,Scientific and Medical(ISM)bands,therefore accessing to this over-crowded wireless medium is usually restricted by international regulations.This is the case of LPWAN solutions such as LoRaWAN or Sigfox,which are subject to a certain duty cycle.Whereas we model the disponibility of the LPWAN interface with the boolean feature D,we consider that the other examined RATs,namely,vehicular WiFi and cellular,do not present this restriction.We also assumethat there is not a dedicated message queue for each RAT.On the contrary,we consider aOne in-One outmessage-forwarding system.We make such assumption given the processing and memory constraints of MCUs,which prevent them of real parallel processing and the capacity of storing great amount of data in RAM memory.Two additional features for defining the message characteristics are also considered in the decision process.The data length (L),in bytes,is an important parameter as some LPWAN technologies have strict limits in the maximum payload size supported in their transmissions.Finally,the urgency level of the message (U) is also taken into account,aiming at considering different types of services,e.g.,high priority: fall detection; medium priority: traffic information; low priority: tracking data.As can be observed,all the considered parameters can be easily collected or computed by the OBU(Figure 2)in a real situation.

Therefore,the system status (S) is defined by the following vector of features(B,{c1,c2,...,cn},D,L,U).This vector should be employed for determining the most adequate action A∈ {a0,a1,...,an},which represents the event of dropping a message (a0) or sending it using one of the available RATs{a1,a2,...,an}.We tackle this issue as a classification problem.Thus,as explained in next sections,our approach considers the DSS to be developed as a supervised multi-class classifier and,to this end,we evaluate a number of well-known specific classification algorithms.Furthermore,having an adequate data set is also crucial for training and evaluating these algorithms.In the following,we provide insights regarding the employed data set.

4.2 Data Set

A synthetic training data set of 10.000 samples,i.e.,features vectors,has been generated by randomly assigning values to each of the input features.The coverage level of each RAT(ci)as well as the packet urgency (U) values have been generated by assigning them random values in the range[0%,100%],following a uniform distribution.Similar procedure has been followed for calculating the values of the OBU’s remaining battery(B)and the packet length(L)features,in this case in the ranges [1%,100%]and [50 bytes,500 bytes],respectively.Finally,the LPWAN interface disponibility(D),modeled as a boolean parameter,has been assigned with true/false values uniformly distributed.Note that this synthetic data set captures the characteristics of realistic situations as all the possible parameter combinations are considered given its size.

The data set has been labelled with the desired action for each of the 10.000 generated samples.For that,we have fixed some hard restrictions in the input features in order to properly select a RAT for sending a given message(Table 1).These have been set considering the power consumption,coverage needs,packet length,and delay constraints of each technology,assigning meaningful values when compared with each other.We have developed a labelling algorithm accounting for these restrictions.After the first round of automatic data labeling,a subsequent manual tuning has been carried out to solve conflicts or ambiguities regarding the final selection,specially focusing on high urgency packets.The labeling processes,both automatic and manual,have been conduced by taking into account the main characteristics of the considered transmission technologies,which are given in Table 2.Therefore,by inspecting tables 1 and 2,the criteria for selecting one RAT or another was as follows.LPWAN is preferred for low-urgency short packets,given its low power consumption,but it is not suitable to transport other type of data packets,due to its limited data-rate and the great latency introduced.In turn,vehicular WiFi technology is devoted to transport highly urgent messages,given its low latency.Finally,given the great capacity of cellular systems,this RAT is employed for transmitting the rest of messages,although with certain restrictions due to its high power consumption.

Table 2.Main characteristics of RATs under consideration.

As a reference,Figure 3 shows the result of labeling 5.000 samples from the training set and gives insights about the effect of packet urgency(U)and battery level(B)features.Each color corresponds to one of the possible actions,namely,select the cellular (green),LPWAN(orange),or Veh.WiFi(blue)interfaces to transmit the message or drop it(black).Observe that most of the dropped packet occurrences happen with low urgency messages specially in the case of low level of battery,although this is also detected in medium/high urgency packets.Those dropped packets with high urgency when the battery level is not critically low are justified by low coverage levels of both vehicular WiFi and cellular technologies.Regarding the usage of each RAT,as expected,vehicular WiFi is employed for packets with high urgency,which is the opposite behaviour in comparison with LPWAN.Finally,cellular is employed for both medium and high urgency packets,but only when a sufficient level of battery is available.

4.3 DSS Implementation

Once generated the data set,a range of multi-class supervised classification algorithms have been trained and their accuracy have been evaluated.Concretely,we have selected the following algorithms: Naive Bayes,multi-layer perceptron,decision tree,and k-Nearest Neighbors (k-NN),given their suitability to be implemented in MCU platforms [22,23].The ML workbench employed for training and evaluating these algorithms was Knime v4.1.2 [24].This analytics platform permits to deploy complex data mining work flows in a graphical way,as shown in Figure 4.In the developed work flow,firstly,the generated data set of 10.000 samples is imported from a text file into the work space.Thereafter,by using stratified sampling,these data are partitioned onto two separated sets: training(75%)and testing(25%)data.The former is used as input for each of the learning algorithms,which produces the corresponding trained model.This output,together with the test portion of data,feeds the associated predictors that returns a final table with the classified data.This table is finally employed by the score modules to evaluate the performance of the different algorithms.Other functions such as statistical analysis or graphical visualization(blue boxes in Figure 4) can be also deployed along the work flow,in order to have a better understanding of the different process’steps.

Figure 4.DSS workbench in Knime.

The four studied classification algorithms have been configured considering the processing and memory constraints posed by typical MCUs such as Arduino Uno.This platform is considered a good testbench,given its wide adoption and hardware constraints.Thus,ML algorithm’s configurations suggested for this board by MCU-specialized libraries [22,23]are employed and,under this limited range of options,we have selected the best parameters to maximize accuracy in each case.The configuration set-up is shown in Table 3.With these settings,the performance of the four algorithms has been evaluated,presenting the main results in next section.

Table 3.Multi-class classification algorithms’ configurations.

V.RESULTS

Table 4 presents the accuracy results for the classification algorithms under consideration.We show the attained accuracy,understood as the number of correct decisions over the number of input samples,and the Cohen’s kappa coefficient,which represents the degree of accuracy and reliability in the classification task.As observed,the best result in terms of accuracy is attained by the decision tree algorithm,with an outstanding performance of 99.5%.Then,k-NN and multi-layer perceptron surpass the 75% of accuracy and,finally,Naive Bayes achieve a performance of 74.6%.These outcomes are confirmed by the obtained Cohen’s kappa coefficient.While Naive Bayes,multilayer perceptron,and k-NN models present a substantial strength of reliability,the decision tree one exhibits an almost perfect classification agreement with the categories assigned in the data set(see[25]).Besides the accuracy and reliability obtained from the classifiers,it is also interesting to evaluate the implications of integrating the DSS within the OBU by considering the system load in terms of energy consumption and QoS.

Table 4.Accuracy results.

Figure 5 presents the distribution of the actions(A)decided by each algorithm.It depicts how the available RATs are employed by each of the investigated algorithms when processing the whole testing data set(2.500 samples).The energy involved by both taking the decision and using each communication technology is also represented.This has been simulated by assigning a fixed transmission power consumption(C)and data-rate (R) to each of the RATs (see Table 5).It can be observed that all the evaluated algorithms present similar levels of energy consumption.However,whereas Naive Bayes,multi-layer perceptron and k-NN make use of more than 55 J,the consumption given by the decision tree algorithm is slightly lower.It can be explained by the fact that the latter is more efficient than the others as the LPWAN technology is used more frequently.Another reason that justifies this outcome is the greater number of packet drops.Nevertheless,as shown in Figure 6,this algorithm makes better choices when dropping packets in comparison with Naive Bayes and k-NN.Observe how both of them present greater numbers of high urgency message dropped,which is completely undesirable.The best result in this regard is attained by the multi-layer perceptron.As explained above,the decision of dropping an urgent message is due to low battery level or scarce coverage of both vehicular WiFi and cellular technologies.

Figure 5.RAT and dropping decisions for packets with respect to their associated energy consumption.

Table 5.RAT characterization.

Figure 6.Impact of packet urgency of dropping decision.

In the light of these outcomes,specially the accuracy results shown in Table 4,and given its greater simplicity to be implemented,the decision tree algorithm outstands as the most convenient one among the studied alternatives.Figure 7 provides further insights regarding its performance in terms of message urgency (U),packet length (L) and RATs’ coverage levels(c1,c2,andc3).As observed in Figure 7a,the favorite RAT for transmitting low urgency short packets is LPWAN;in turn,medium and high urgency packets are preferred to be sent through cellular and vehicular WiFi technologies,respectively.The greatest amount of dropped packets are those with low urgency and big size.Figure 7b,Figure 7c and Figure 7d reveal interesting performance trends considering the coverage level of each technology.Figure 7b shows how the coverage level of LPWAN is not a barrier for its selection,given the great sensitivity provided by this RAT.Figure 7c shows how urgent packets are transmitted through the cellular transceiver with low vehicular WiFi coverage levels.Cellular technology may be understood as a backup technology,specially when highly urgent messages cannot be sent through the preferred RAT(vehicular WiFi in this case).As observed in Figure 7d,medium priority messages are the most affected ones when the cellular coverage is poor,as just some of them are transmitted by LPWAN and the rest are dropped.

Figure 7.Actions chosen by the computed decision tree in terms of(c1,c2,c3,L,U)input features.

Finally,as mentioned above,the resulting decision tree has been implemented in an Arduino Uno board for evaluating its performance.The developed tree has nine levels and consists of 28 nodes and 29 leaves.Given its size,it has not been reproduced here.Once coded in the Arduino Uno,the DSS function just takes 444 bytes from program memory,hence leaving free 99%of this scarce resource.For its evaluation we have introduced 1.000 samples,which has been individually classified by the implemented function.The performance of this real implementation has been benchmarked by measuring the needed time by an Arduino Uno to process it.The decision time took,in average for the 1.000 introduced samples,51.36±1.09µs(confidence intervals withα=0.05).A sample of 100 observations is presented in Figure 8.These results,together with its low memory footprint showed previously,support the suitability of this algorithm to be implemented in constrained processing units.Therefore,we can conclude that introducing an intelligent DSS for multi-RAT decision in IoMT devices is a good choice to obtain good overall classification results while maintaining a low energy profile.

Figure 8.Processing time of the implemented decision tree model running on an Arduino Uno.

VI.CONCLUSIONS

The arrival of the IoMT has paved the way for the development of personalized and dynamic services to end-users that may seamlessly consume them across different intelligent environments.In this line,citizens are adopting new urban transportation models by the use of personal mobility vehicles such as shared bikes,segways or electric scooters.In order to exploit efficiently all the potential that hyper-connected smart cities and C-ITS scenarios bring,these personal mobility devices should incorporate OBUs with a range of RATs,e.g.,cellular,LPWAN,vehicular WiFi,etc.However,given the stringent energy constraints of IoMT elements,it is crucial to provide them with intelligent algorithms aiming at selecting the most adequate RAT to be employed considering the OBU’s status as well as the characteristics of the message to be sent.In this work we have provided a conceptual multi-RAT OBU architecture for eco-friendly personal mobility devices.Then it has been provided with a DSS solving the RAT selection problem.To this end,a number of ML supervised classification algorithms,namely,Naive Bayes,multi-layer perceptron,decision tree and k-NN,have been explored.From this study,the decision tree algorithm has been selected as the most efficient one in terms of QoS,complexity and energy consumption; hence,it has been implemented in a real MCU.From the attained results,it has been demonstrated the good performance of the DSS,which permits to save energy as well as to deal with the urgency required by certain transmissions.The real implementation for an Arduino Uno platform requires few program memory,444 bytes,and it can be executed in around 50µs.As future work,it is planned to continue exploring the field of ML for improving IoMT operations.Concretely,we plan to evaluate the performance of other ML algorithms as well as considering other input features such as the current traffic load supported by each interface.Besides,we plan to build a real prototype for validating our proposal in a constrained device with real data in terms of localdata processing performance and energy savings using wireless communications.At the same time,a simulation framework is being developed to better evaluate algorithm variations at earlier stages before real testing.

ACKNOWLEDGEMENT

This work has been supported by the Spanish Ministry of Science,Innovation and Universities,under the Ramon y Cajal Program (ref.RYC-2017-23823)and the projects PERSEIDES (ref.TIN2017-86885-R) and Go2Edge (ref.RED2018-102585-T); by the European Commission,under the 5G-MOBIX(Grant No.825496) and IoTCrawler (Grant No.779852)projects; and by the Spanish Ministry of Energy,through the project MECANO (ref.PGE-MOVESSING-2019-000104).