Multi-Step Reasoning for IoT Devices

Jos

e Miguel Blanco

1 a

and Bruno Rossi

2 b

Escuela T

ecnica Superior de Ingenieros de Telecomunicaci

on, Universidad Polit

ecnica de Madrid, Spain

Faculty of Informatics, Masaryk University, Brno, Czech Republic

Keywords:

Internet of Things (IoT), Web of Things (WoT), Semantic Reasoning.

Abstract:

Internet of Things (IoT) devices are constantly growing in numbers, forecasted to reach 27 billion in 2025.

With such a large number of connected devices, energy consumption concerns are a major priority for the

upcoming years. Cloud / Edge / Fog Computing are critically associated with IoT devices as enablers for data

communication and coordination among devices. In this paper, we look at the distribution of Semantic Rea-

soning between IoT devices and deﬁne a new class of reasoning, multi-step reasoning, that can be associated at

the level of the edge or fog node in the context of IoT devices. We conduct an experiment based on synthetic

datasets to evaluate the performance of multi-step reasoning in terms of power consumption, memory, and

CPU usage. Overall we found that multi-step reasoning can help in reducing computation time and energy

consumption on IoT devices in the presence of larger datasets.

1 INTRODUCTION

Internet of Things (IoT) refers to a large number of

physical devices, sensors, software artifacts, all with

processing capabilities that can interact over the more

disparate communication network to provide smart

services to users (Dorsemaine et al., 2015; Atzori

et al., 2017). The number of IoT devices deployed

globally in the world has constantly been growing,

increasing 9% in 2021, representing 12.3 billion con-

nected devices with 27 billion IoT devices forecasted

for 2025 (Hasan, 2021).

Natural support to IoT was the emergence of Edge

and Fog Computing – bringing the computation and

data storage closer to the devices that are generat-

ing, interacting, and integrating data (Ai et al., 2018;

Satyanarayanan, 2017). As such, with Edge Comput-

ing, the data is processed directly on the device/sen-

sor at the edge without any data transfer, while in Fog

Computing, data can be processed in an intermediary

node close to the edge, differently to Cloud Comput-

ing in which data needs to be propagated back to the

Cloud for processing (Ai et al., 2018). The advantage

in the context of IoT is that the large amount of data

generated by different IoT devices can be processed

directly at the edge of the network – on the other hand,

more processing power is required (Ai et al., 2018).

https://orcid.org/0000-0001-9460-8540

https://orcid.org/0000-0002-8659-1520

While Edge Computing can bring services with faster

response time, reasoning at the level of edge devices

still needs to take into account the knowledge gen-

erated by peer devices to grant Semantic Reasoning

based on the knowledge derived from all the con-

nected IoT devices (Jara et al., 2014). Furthermore,

energy consumption patterns need to be considered

when moving reasoning from the Cloud to edge de-

vices (Mocnej et al., 2018; Cui et al., 2018).

In this paper, we look at the concept of Semantic

Reasoning in distributed and connected IoT devices.

In particular, we look at how different distributions of

reasoning at divergent moments in time can inﬂuence

computational performance and energy consumption

during the reasoning process. We set up an experi-

ment with synthetic datasets looking at different rea-

soning rules and evaluating the performance of alter-

native reasoning types.

We have the following contributions in this paper:

1. Deﬁnition of the concept of multi-step reasoning

for Semantic Reasoning of IoT devices at the Fog

or Edge level;

2. An experiment with synthetic generated data,

studying the different properties of multi-step rea-

soning – the experimental package is available on

Figshare (see Section 5);

The paper is structured as follows. In Section 2 we

review the background about IoT, Edge and Fog Com-

330

Blanco, J. and Rossi, B.

Multi-Step Reasoning for IoT Devices.

DOI: 10.5220/0011772700003464

In Proceedings of the 18th International Conference on Evaluation of Novel Approaches to Software Engineering (ENASE 2023), pages 330-337

ISBN: 978-989-758-647-7; ISSN: 2184-4895

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

puting and the connection with Semantic Reasoning.

In Section 3 we provide the fundamental deﬁnitions

of multi-step reasoning. In Section 4 we provide a

scenario in the context of a smart energy domain with

charging station, smart homes, and electric vehicles

– all needing reasoning capabilities for the most efﬁ-

cient utilization of resources. In Section 5 we delve

into the experimental design, explaining how the fun-

damentals of multi-step reasoning were instantiated in

terms of implementation and the datasets used for the

evaluation of computational performance and energy

consumption. In Section 6 we provide the main re-

sults from the experiment together with the threats to

validity. In Section 7 we provide a discussion about

the result in relation to the proposed research ques-

tions. Section 8 concludes the paper.

2 BACKGROUND

The emergence of Cloud Computing was one of the

cornerstones for the evolution and growth of IoT

(Satyanarayanan, 2017). Initially, IoT devices would

represent devices with limited resources, thus with

reduced possibility of in-place reasoning algorithms

implementations and data aggregation (Lea, 2018).

However, as the added value of IoT devices is in the

aggregation and integration of data from heteroge-

neous and numerous devices (such as thousands of

sensors, cameras, smartphones, etc...), the Cloud con-

stituted the ﬁrst step towards added capabilities — in

terms of data ingestion, aggregation, analytics, ma-

chine learning, among others (Fig. 1).

As such, Edge Computing, placing computation

at the edge of the network in proximity to mobile de-

vices and sensors, emerged as a way to reduce latency

issues and move the data processing part closer to

the devices themselves (Satyanarayanan, 2017). Con-

versely, Fog Computing emerged as an extension to

the Cloud to support devices with a layer between the

edge and the Cloud (Fig. 2). Both Edge / Fog Com-

puting have advantages over traditional IoT Cloud-

based topologies: more responsive services, support

of edge analytics for scalability, enforcement of pri-

vacy policies, or the possibility to mask Cloud out-

ages (Satyanarayanan, 2017).

However, one critical turning point for the deci-

sion about the network topology is that communica-

tion from the Cloud infrastructure to edge gateway

and edge devices is subject to disparate communica-

tion latency: from the real-time performance at the

level of smart devices to milliseconds latency at the

edge gateway and potentially milliseconds to seconds

at the level of the Cloud (Lea, 2018).

Furthermore, recently Green Computing concerns

have emerged: green computing has been deﬁned

as ”the environmentally responsible and eco-friendly

use of computers and their resources” (Salama,

2020). With the increasing number of connected

devices and computational units, the goal of efﬁ-

cient power consumption has become a key target for

the IT domain, making considerations about power

consumption one of the key aspects of network and

computation optimizations— leading to the so-called

Green IoT (Zhang et al., 2018; Lyu et al., 2018). This

brings considerations about the power consumption

of computations and communications related to edge

devices (Mocnej et al., 2018; Cui et al., 2018).

Additionally, IoT devices have become pervasive

in most of our everyday life aspects: from grant-

ing smart sensors to providing remote access to dif-

ferent systems (Arfaoui et al., 2020). Nevertheless,

this has also prompted some problems regarding the

interoperability of these devices. One of the key

points for the extended connectivity of IoT devices,

and solving the interoperability problem, is their in-

teraction with the Semantic Web through the stan-

dard known as the Web of Things (WoT). WoT pro-

vides standardized metadata and other re-usable tech-

nological building blocks to enable easy integration

across IoT platforms and application domains

. This

allows for processing the generated data through Se-

mantic Reasoning to power the inference capabili-

ties of devices (Kisliuk et al., 2022) or increase the

security for knowledge bases queries (Krishnasamy-

Sivaprakasam and Slutzki, 2021). In particular, the

extension of IoT into WoT allows for developing a

formal space with all the advantages that can be ex-

tracted from it.

The techniques used to extract knowledge from

the data generated by IoT devices, such as sensors,

is by semantic reasoning (Maarala et al., 2017). This

method is based on the use of logical rules to de-

rive conclusions from the annotations that the se-

mantic web framework has provided (Hitzler and

Van Harmelen, 2010). These rules, applied by us-

ing a semantic web reasoner and sometimes strength-

ened by the structure provided by an ontology, help

to present the user with an extended representation of

the situation so any measures can be adopted. In par-

ticular, it is necessary to mention that the deployment

of the reasoner and rules, with regards to the architec-

ture, its absolutely critical in time constrained situa-

tions. For that matter some effort has been made to

ensure that the required tasks are performed as fast as

possible (Wang et al., 2018).

https://www.w3.org/WoT/

Multi-Step Reasoning for IoT Devices

331

Figure 1: Cloud, Edge Gateway, and Smart Devices (adapted from (Lea, 2018)).

3 FUNDAMENTALS

In this section, we detail the main fundamentals about

Semantic Reasoning in the context of IoT devices and

introduce the concept of multi-step reasoning.

Deﬁnition 1 (Reasoning). Reasoning is the automatic

procedure to generate new knowledge and relation-

ships based on provided data and a set of additional

rules. Generally speaking, this procedure executes

the schema provided by the rules to obtain new data

at once.

Deﬁnition 2 (Multi-step reasoning). Multi-step rea-

soning is a method that acts as a subclass of Reason-

ing as deﬁned above. This subclass is characterized

by executing the rules’ schema one at a time instead

of all of them at once.

Now, we will consider the following rules:

Figure 2: Fog Computing and smart devices (adapted from

(Lea, 2018)).

R1. [ruleConjunction: (?a :and ?b) (?b :and ?c) → (?a

:and ?c)]

R2. [ruleTransitivity: (?a :then ?b) (?b :then ?c) → (?a

:then ?c)]

Both rules share the same structure: they have two

different triples in the ﬁrst term that share the same

predicate, :and and :then, and both have as a conclu-

sion one triple in which elements of both of the previ-

ous triples are combined. This means that the rules are

comparable in terms of depth of the triples, elements

of the terms and general structure. One might argue

that both rules are indeed just different instances of

the same general rule. A general rule that could be

read as follows:

[ruleGeneral (?a :predicate ?b) (?b :predicate ?c)

→ (?a :predicate ?c)]

And while the argument holds some value, it is nec-

essary to understand that these rules have not been

selected by their compliance with the Resource De-

scription Framework (RDF), one of the most com-

mon building blocks of the semantic web, structure,

but rather for their logical value. Because of this, we

could understand the rules as Hilbert-style rules for-

malized as follows:

R1. (A ∧ B) & (B ∧ C) ⇒ (A ∧ C)

R2. (A → B) & (B → C) ⇒ (A → C)

As we can see, there are no common connectives be-

tween the two rules aside from the meta-connectives

& and ⇒. Therefore, if we are deﬁning both con-

nectives of the rules, conjunction (∧) and conditional

(→) as primitives from a certain algebra, we should

be in the clear as to why both rules are of interest and

cannot be inferred one from the other.

ENASE 2023 - 18th International Conference on Evaluation of Novel Approaches to Software Engineering

332

4 IoT SCENARIO

To showcase the concept of multi-step reasoning in

the cyber-physical context, we can provide a scenario

with Smart Meters, data aggregators/concentrators,

and central servers in a typical smart energy applica-

tion (e.g., (Chren et al., 2016)), for which there has al-

ready been semantic work done (Blanco et al., 2023).

An example of the proposed network is the one

established by a Smart City in which we would have

smart devices and smart sensors at the edge that re-

quire total availability to capture the multiple streams

of data in real-time (e.g., (Schleicher et al., 2016)).

On the other hand, the data center acting as a cen-

tral server would be used to process the data received

from the edge devices, ensure that all the services are

working properly, and make possible predictions on

future events and requirements. The ﬁnal third ele-

ment is a data aggregation point. This point would

collect the data generated by the smart devices and

smart sensors and pre-process it so the work that

needs to be done in the central server would be less

and could be carried out in a more manageable and

efﬁcient way.

This network would go on to generate certain data

in the shape of triples. Some of these triples could be

the following:

T1. :smartcar :in :chargestation

T2. :chargestation :in :mainstreet

T3. :smartcar :charges :batterycar

T4. :batterycar :charges :1.5Kw/h

The ﬁrst two triples, T1 and T2, could be processed

with the ruleConjunction that we have included be-

fore. In the same way, T3 and T4 could be processed

with ruleTransitivity. The outcomes would be, re-

spectively, as follows:

T5. :smartcar :in :mainstreet

T6. :smartcar :charges :1.5Kw/h

This would allow the main data center of the Smart

City to know about the situation of the car as well

as the task that is being performed. In this case, we

would be able to infer that the car is being charged

at a rate of 1.5Kw/h on the main street. Informa-

tion that should be fundamental when considering the

energy required by a certain part of the city and the

trafﬁc management of said part of the city. This ob-

viously is not a problem when dealing with just one

element (in this case, the smart car) but could be a po-

tential overload when dealing with thousands of vehi-

cles and the streams of data that are constantly gen-

erated. Therefore, it is necessary to ﬁnd a way to

improve the performance of reasoning when dealing

with datasets that can contain thousands, if not mil-

lions, of elements.

5 EXPERIMENTAL EVALUATION

We run an experiment with the following goal: to

evaluate the multi-step reasoning (Deﬁnition III.2) in

terms of power consumption and computational per-

formance of IoT devices when compared with tradi-

tional reasoning (Deﬁnition III.1).

Graphs of the results discussed in this and further

sections and all the models and experiment’s scripts

are available in a Figshare repository (Blanco and

Rossi, 2023)

where the reader might consult them

at any given time.

For multi-step reasoning, we adopted Apache Jena

and used N3 ﬁles. In a production environment, the

usual ﬂow is to load the rules after one reasoning

model is created. The traditional technique for im-

plementing a reasoning model (Def III.1) consists of

creating the model and loading it with a set of rules

that encompasses all the rules that are to be used or

that are plainly considered, even if they are not used.

In the multi-step reasoning paradigm we are propos-

ing (Def III.2), this will happen as different instances

of loading sets of rules, one for each rule. The load-

ing of these sets excludes the rules that are not used in

further pursuit of efﬁciency and performance as, even

if it is just a fraction of a second, some time, energy,

and other resources would be saved.

5.1 Research Questions

We have the following Research Questions (RQs):

Research Questions

RQ1. What is the energy consumption when the

reasoning is done as multi-step reasoning

compared when it is done traditionally?

RQ2. Can the performance of the reasoning

(time and memory / CPU usage) be im-

proved when it goes through the means of

multi-step reasoning as opposed to the tra-

ditional technique?

5.2 Datasets

For the experimental evaluation, we used datasets that

have been synthetically generated. These datasets

https://doi.org/10.6084/m9.ﬁgshare.19493996.v1

Multi-Step Reasoning for IoT Devices

333

have been created sequentially with a number n, for

n ∈ N, of elements per rule thanks to a Python script.

The datasets contain an increasing number of cases

that are constructed in such a way that the rules in-

troduced are forced to process all the elements of the

datasets and all the outcomes. This means that, at the

end of the reasoning, the reasoner would deal with a

relationship between the ﬁrst and last elements. An

excerpt of any of these datasets is as follows:

...

:n-1 :and :n-2

:n-1 :then :n-2

:n :and :n-1

:n :then :n-1

These datasets have been designed so there is no

data overload on the size of the triples, but rather

can test the performance of the rules engine. Fur-

thermore, they can provide a scenario in which the

model is asked to perform a scalable number of

tasks, thus showing the performance of the reasoner

against multiple, totally different scenarios. The dif-

ferent datasets are, by name and total number of el-

ements: DS1(20), DS2(200), DS3(500), DS4(1000),

and DS5(2000).

Table 1: Average Energy Usage.

Traditional Reasoning (J) Multi-Step (J)

DS1 - CPU cores 48.74 56.48

DS2 - CPU cores 127.17 122.71

DS3 - CPU cores 723.02 520.65

DS4 - CPU cores 9 723.97 6 493.61

DS5 - CPU cores 157 939.74 105 357.26

DS1 - RAM 17.53 18.20

DS2 - RAM 24.49 24.95

DS3 - RAM 125.57 90.51

DS4 - RAM 1 659.19 1 084.57

DS5 - RAM 32 385.26 18 921.56

DS1 - PKG 114.07 126.12

DS2 - PKG 214.90 213.58

DS3 - PKG 1 209.13 866.76

DS4 - PKG 16 616.42 10 862.94

DS5 - PKG 282 784.75 183 546.67

5.3 Analysis Method

The energy usage has been measured using the perf

tool and is in Joules (J). The machine used is a In-

tel CPU Core i7-4790@3.60GHz, 16Gb RAM DDR3

@ 1600 MHz, and running Ubuntu 20.04.3 LTS. The

IDE utilized for the implementation and testing is In-

telliJ Idea Ultimate 2021.2. Each benchmark was per-

formed ﬁve (5) times, considering the average. Fi-

nally, the reasoning engine was implemented using

Apache Jena 4.4.0.

Table 2: Average Measurements for Traditional Reasoning.

Time (ms) Max Memory (KB) Max CPU (%)

DS1 256 13 054 6.96

DS2 1 833 120 450 39.10

DS3 40 178 157 445 32.94

DS4 667 629 222 322 33.16

DS5 13 025 579 353 993 34.90

6 EXPERIMENTAL RESULTS

The initial measurements, performed with the small-

est dataset, DS1 of 20 elements, show that the per-

formance of traditional reasoning is better than the

new alternative introduced by multi-step reasoning

(481ms vs 225.9ms on average, Tables 2,3). The

results show an increase of almost a 100% in time

to perform the reasoning when comparing the multi-

step reasoning to the traditional way. Despite that,

the maximum memory and CPU usage are similar.

In a similar sense, only the energy consumption of

the CPU cores is above a 15% in multi-step reason-

ing, while the energy consumption of the RAM and

the PKG remains stable all through the different tech-

niques (Table 1).

Table 3: Average Measurements for Multi-Step Reasoning.

Time (ms) Max Memory (KB) Max CPU (%)

DS1 481 12 902 6.56

DS2 1 942 30 484 34.38

DS3 26 543 150 845 33.63

DS4 403 304 190 399 31.78

DS5 7 358 757 245 303 38.4

Similarly, the results for DS2, of 200 elements,

are comparable in time spent and Maximum CPU us-

age. Also, the energy consumption across CPU cores,

RAM, and PKG is almost the same in traditional and

multi-step reasoning (Table 1). The main difference

resides in the max memory usage, where traditional

reasoning requires almost four (4) times the amount

that multi-step reasoning does (Tables 2,3).

By the time that we reach DS3, of 500 elements,

the multi-step reasoning starts to show improvements.

Time spent reasoning is almost cut in half by multi-

step reasoning in comparison to traditional methods

(26 543ms vs 40 178ms, Tables 2,3). Also, the en-

ergy consumption of the CPU cores, the RAM and the

PKG in traditional reasoning are increased by a third

once compared to multi-step reasoning (Table 1). One

trend that we see at this point and is maintained in fur-

ther datasets is that the Maximum Memory and CPU

usage are similar in traditional and multi-step reason-

ing. Nevertheless, the overall CPU usage is lower in

multi-step reasoning with lower spikes over time.

ENASE 2023 - 18th International Conference on Evaluation of Novel Approaches to Software Engineering

334

100

200

300

400

500

Average Time (in ms)

Traditional Reasoning

Multi-Step Reasoning

(a) DS1(10)

500

1,000

1,500

2,000

Average Time (in ms)

(b) DS2(100)

·10

Average Time (in ms)

0.5

2.5

4.5

6.5

·10

Average Time (in ms)

(d) DS4(500)

0.3

0.6

0.9

1.2

·10

Average Time (in ms)

(e) DS5(1000)

Figure 3: Average execution times of different datasets (note different scale).

The most signiﬁcant differences are perceived in

the test runs on DS4 and DS5. DS4, of 1000 elements,

shows that the energy consumption of traditional rea-

soning is 1.5 times bigger than that of multi-step rea-

soning in all the measurements made: CPU cores,

RAM and PKG (Table 1). Both maximum memory

and CPU usage are similar as we pointed in the pre-

vious paragraph. Nevertheless, the time of reason-

ing, when done traditionally, is more than 1.5 longer,

which, at this point, accounts for more than 4 min-

utes of difference (404 304ms (6.7min) vs 667 629ms

(11.1min), Tables 2,3).

In case of DS5 of 2000 elements, the energy con-

sumption almost doubles in PKG and RAM, while it

is up to a 1.5 times more than multi-step reasoning

in CPU cores (Table 1). Both maximum memory and

CPU use is comparable in both approaches. However,

the time spent reasoning is where we see a difference

of double the amount in traditional against multi-step

reasoning. If in DS4 the time difference accounted

for 4 minutes of difference, here we have a differ-

ence of up to two hours on the averages (7 358 757ms

(∼122min) vs 13 025 579ms (∼217min), Tables 2,3).

6.1 Threats to Validity

We have some threats to validity of the experiment.

First of all, in this paper, we deal with the rea-

soning that is performed on the edge device or at the

fog node by looking at the different reasoners’ rules

that can be involved. We do not deal with the net-

work aspects, such as delays and transmission power

requests, that can have a relevant impact. Other re-

search has dealt with measuring an extra amount of

energy and time required for the transmission and

reception of tasks dived among edge and fog nodes

– so-called ofﬂoading (Bozorgchenani et al., 2018).

Having a view that includes Semantic Reasoning and

networking considerations can be considered the next

step for research.

The experiment has only been performed by look-

ing at two different rules. The results might vary when

the set of rules reaches a certain size or that, after a

certain point, we could see a diminishing returns sit-

uation. We also considered synthetic data generated

according to typical scenarios. We plan to extend the

evaluation to the reasoning on IoT devices.

Furthermore, we can point out that our testing was

done utilizing Apache Jena’s reasoner. Other rea-

soning models might perform differently. Neverthe-

less, this should be a limited concern as the most ex-

tended and adopted reasoning model is the one of

Apache Jena. The conﬁguration details and experi-

mental package are available in Section 5.

Additionally, we performed the experiment con-

sidering abstract rules instead of an ontology. We

intended to test with synthetically generated datasets

with the idea that techniques should be ﬁrst tried out

in an in-vitro environment before including other vari-

ables such as ontologies, as those can affect perfor-

Multi-Step Reasoning for IoT Devices

335

mance as seen in (Blanco et al., 2021). We plan to

adopt more complex models in future works.

7 DISCUSSION

As the results have shown, the traditional way of rea-

soning has a higher performance, and lower energy

consumption than multi-step reasoning, whenever the

datasets are below 100-200 elements for each rule.

This is related to the time needed to perform the rea-

soning, as traditional reasoning takes reduced time

compared to multi-step reasoning.

Nevertheless, this trend changes whenever the

datasets get larger. Once the barrier of 100 elements

per rule is passed – in the case of this experiment,

once the 500 elements dataset has been reached – the

performance, time spent for reasoning, and energy

consumption are in favour of multi-step reasoning.

At ﬁrst, it is a discreet improvement, with a differ-

ence of only mere seconds. However, once the dataset

reaches the size of 2000 elements, we see a consistent

improvement.

RQ1 Findings

¬ Data collected in Table 1, shows that multi-

step reasoning is more energy efﬁcient way to

process data. The disparity between energy

consumption gets bigger with the increasing

cardinality of the datasets.

Although for the ﬁrst dataset multi-step reasoning

shows that it is not the best option for small datasets,

results improve on the other datasets with larger

number of elements. This propels multi-step reason-

ing as a more energy efﬁcient option than traditional

reasoning, especially under large workloads.

RQ2 Findings

¬ The performance in terms of memory/CPU

usage is better when addressed from a tradi-

tional perspective as long as the datasets re-

main small (up to 100 elements). From this

point on, the introduction of multi-step rea-

soning improves the performance, in terms of

memory/CPU and time spent (Tables 1,2,3).

About RQ2, it is worth noting that CPU does not

see a difference as the one that can be noted in time

and energy consumption. Particularly, if we strictly

speak about maximum numbers, we can see that the

max CPU usage in multi-step reasoning is higher than

the traditional one, but this is remedied by a differ-

ence in the usage of the CPU. We can also note differ-

ent patterns in CPU usage comparing multi-step and

traditional reasoning

with more spikes in % of CPU

usage. The rest of the data shows that multi-step rea-

soning is much more efﬁcient in both time needed

and max memory utilized. While the difference in

max memory might not be signiﬁcant, the difference

in time is more prominent. This can be seen as the re-

duction in time needed to compute the biggest dataset

is almost double, with a difference of almost 2 hours.

This would appear crucial for developing a reasoning

scheme to process data in real-time.

Based on the experimental data, we can suggest

that multi-step reasoning is a better technique than

traditional reasoning, as it improves all facets: time,

max memory, max CPU, and energy consumption.

Given the current state of the art and the amount of

data generated by the billions of online devices, the

performance in larger datasets is more relevant. The

difference in the case of small datasets is minimal:

less than a quarter of a second and less than 20 Joules

overall for datasets of 20 elements and less than one-

tenth of a second in datasets of 200 elements.

8 CONCLUSION

With the importance that Edge and Fog Computing

have acquired in recent years, it is fundamental to sup-

port data processing at the level of IoT devices. One

step in this direction is to provide Edge devices with

Semantic Reasoning capabilities.

In this paper, we have looked at the concept of

Semantic Reasoning in distributed and connected IoT

devices. In particular, we proposed multi-step reason-

ing as an approach to improve the performance of rea-

soning and conducted an experiment with synthetic

data to evaluate IoT devices’ computational perfor-

mance and energy consumption.

Overall, we found that multi-step reasoning can

help reduce computation time and energy consump-

tion in larger datasets. Our in-vitro evaluation will be

further expanded in future works considering the de-

ployment of ontologies. It is also crucial to expand the

work to include more complex rules that intertwine;

this would allow us to showcase the performance of

multi-step reasoning in a situation where the rules

are not considered as stand-alone entities but rather

a complex system that beneﬁts from previous results

to get more robust conclusions.

The ﬁgures are available in the additional material on

Figshare

ENASE 2023 - 18th International Conference on Evaluation of Novel Approaches to Software Engineering

336

ACKNOWLEDGEMENTS

The research was supported from ERDF/ESF ”Cyber-

Security, CyberCrime and Critical Information Infras-

tructures Center of Excellence” (No. CZ.02.1.01/0.0

/0.0/16 019/0000822).

This work has been supported by the Madrid

Government (Comunidad de Madrid-Spain) under the

Multiannual Agreement with Universidad Polit

ecnica

de Madrid in the line Support for R&D projects for

Beatriz Galindo researchers, in the context of the V

PRICIT (Regional Programme of Research and Tech-

nological Innovation).

REFERENCES

Ai, Y., Peng, M., and Zhang, K. (2018). Edge computing

technologies for internet of things: a primer. Digital

Communications and Networks, 4(2).

Arfaoui, A., Cherkaoui, S., Kribeche, A., and Senouci,

S. M. (2020). Context-aware adaptive remote access

for iot applications. IEEE Internet of Things Journal,

7(1).

Atzori, L., Iera, A., and Morabito, G. (2017). Understand-

ing the internet of things: deﬁnition, potentials, and

societal role of a fast evolving paradigm. Ad Hoc Net-

works, 56.

Blanco, J., Rossi, B., and Pitner, T. (2021). A compari-

son of smart grids domain ontologies. In Proceedings

of the 17th International Conference on Web Informa-

tion Systems and Technologies - Volume 1: WEBIST,,

pages 115–123. INSTICC, SciTePress.

Blanco, J. M. and Rossi, B. (2023). Multi-Step Reasoning

for IoT Devices.

Blanco, J. M., Rossi, B., and Pitner, T. (2023). A com-

parative study of energy domain ontologies. In Web

Information Systems and Technologies, pages 43–58.

Springer International Publishing.

Bozorgchenani, A., Tarchi, D., and Corazza, G. E. (2018).

Centralized and distributed architectures for energy

and delay efﬁcient fog network-based edge comput-

ing services. IEEE Transactions on Green Communi-

cations and Networking, 3(1).

Chren, S., Rossi, B., and Pitner, T. (2016). Smart grids de-

ployments within eu projects: The role of smart me-

ters. In 2016 Smart cities symposium Prague (SCSP).

IEEE.

Cui, L., Xu, C., Yang, S., Huang, J. Z., Li, J., Wang, X.,

Ming, Z., and Lu, N. (2018). Joint optimization of

energy consumption and latency in mobile edge com-

puting for internet of things. IEEE Internet of Things

Journal, 6(3).

Dorsemaine, B., Gaulier, J.-P., Wary, J.-P., Kheir, N., and

Urien, P. (2015). Internet of things: a deﬁnition & tax-

onomy. In 2015 9th International Conference on Next

Generation Mobile Applications, Services and Tech-

nologies. IEEE.

Hasan, M. (2021). State of IoT 2021: Number of connected

IoT devices growing 9% to 12.3 billion globally, cel-

lular IoT now surpassing 2 billion.

Hitzler, P. and Van Harmelen, F. (2010). A reasonable se-

mantic web. Semantic Web, 1.

Jara, A. J., Olivieri, A. C., Bocchi, Y., Jung, M., Kastner,

W., and Skarmeta, A. F. (2014). Semantic web of

things: an analysis of the application semantics for the

iot moving towards the iot convergence. International

Journal of Web and Grid Services, 10(2-3).

Kisliuk, B., Tieben, C., Niemann, N., Br

ocker, C., Linge-

mann, K., and Hertzberg, J. (2022). Online Infer-

ence of Robot Navigation Parameters from a Semantic

Map. In 14th International Conference on Agents and

Artiﬁcial Intelligence.

Krishnasamy-Sivaprakasam, G. and Slutzki, G. (2021).

Secrecy-preserving reasoning in acyclic dl-liter

knowledge bases in the presence of bcqs. In Proceed-

ings of the 13th International Conference on Agents

and Artiﬁcial Intelligence - Volume 2: ICAART,.

SciTePress.

Lea, P. (2018). Internet of Things for Architects: Architect-

ing IoT solutions by implementing sensors, communi-

cation infrastructure, edge computing, analytics, and

security. Packt Publishing Ltd.

Lyu, X., Tian, H., Jiang, L., Vinel, A., Maharjan, S., Gjess-

ing, S., and Zhang, Y. (2018). Selective ofﬂoading

in mobile edge computing for the green internet of

things. IEEE network, 32(1).

Maarala, A. I., Su, X., and Riekki, J. (2017). Semantic

reasoning for context-aware internet of things appli-

cations. IEEE Internet of Things Journal, 4(2).

Mocnej, J., Mi

skuf, M., Papcun, P., and Zolotov

a, I. (2018).

Impact of edge computing paradigm on energy con-

sumption in iot. IFAC-PapersOnLine, 51(6).

Salama, M. (2020). Green Computing, a contribution to

save the environment.

Satyanarayanan, M. (2017). The emergence of edge com-

puting. Computer, 50(1).

Schleicher, J. M., V

ogler, M., Dustdar, S., and Inzinger, C.

(2016). Enabling a smart city application ecosystem:

requirements and architectural aspects. IEEE Internet

Computing, 20(2).

Wang, S., Wan, J., Li, D., and Liu, C. (2018). Knowledge

reasoning with semantic data for real-time data pro-

cessing in smart factory. Sensors, 18(2).

Zhang, K., Leng, S., He, Y., Maharjan, S., and Zhang, Y.

(2018). Mobile edge computing and networking for

green and low-latency internet of things. IEEE Com-

munications Magazine, 56(5).

Multi-Step Reasoning for IoT Devices

337