Available online at www.sciencedirect.

com
Available online at www.sciencedirect.com
Available online at www.sciencedirect.com

ScienceDirect
Procedia Computer Science 00 (2019) 000–000
Procedia
Procedia Computer
Computer Science
Science 17000 (2019)
(2020) 000–000
362–369 www.elsevier.com/locate/procedia
www.elsevier.com/locate/procedia

The 11th International Conference on Ambient Systems, Networks and Technologies (ANT)
The 11th International Conference
April on
6 - Ambient Systems,Poland
9, 2020, Warsaw, Networks and Technologies (ANT)
April 6 - 9, 2020, Warsaw, Poland
A
A New
New Workload
Workload Prediction
Prediction Model
Model Using
Using Extreme
Extreme Learning
Learning
Machine and Enhanced Tug of War optimization
Machine and Enhanced Tug of War optimization
Thieu Nguyenaa , Bao Hoangaa , Giang Nguyenbb , Binh Minh Nguyena,∗
Thieu Nguyen , Bao Hoang , Giang Nguyen , Binh Minh Nguyena,∗
a School of Information and Communication Technology, Hanoi University of Science and Technology, Hanoi, Vietnam
a School of Information and Communication Technology, Hanoi University
b Institute of Informatics, Slovak Academy of Sciences,ofBratislava,
Science and Technology, Hanoi, Vietnam
Slovakia
b Institute of Informatics, Slovak Academy of Sciences, Bratislava, Slovakia

Abstract
Abstract
Time series data is widely accessible in many life areas like economy, weather, stock price, retail sales, distributed system work-
Time
loads.series
Whiledata
manyis studies
widely have
accessible
focusedin on
many life areas
improving like economy,
existing predictionweather,
techniquesstock
onprice, retail
accuracy sales,less
aspect, distributed
efforts issystem work-
paid towards
loads. While many studies have focused on improving existing prediction techniques on accuracy aspect, less efforts
simple but efficient forecasting models in order to keep the balance between computation cost and prediction accuracy. In this work, is paid towards
simple but efficient
we propose a novel forecasting
time series models in order
prediction model, to which
keep the balance
aims between
to both modelcomputation
simplicity andcost and prediction
accuracy. The coreaccuracy. In thisiswork,
of the model built
we propose
based a novel
on extreme time series
learning prediction
machine. Due tomodel, whichdetermination
the random aims to both model
processsimplicity and accuracy.
for input weights The core
and hidden of the
biases, modellearning
extreme is built
based
machineon requires
extreme learning machine.
a large number of Due
hiddento the random
neurons determination
to achieve processand
good results forthis
input weightsthe
increases andmodel
hiddencomplexity.
biases, extreme learning
To overcome
machine requires
this drawback, wea propose
large number
a newofopposition-based
hidden neurons to tugachieve
of wargood results and
optimization this increases
to select optimallytheinput
model complexity.
weights, To overcome
and hidden biases
this
then drawback, we propose
apply to extreme a new
learning opposition-based
machine. Two real publictug of warmonitoring
traffic optimization to select
datasets optimally
from Internet input weights,
service andwere
providers hidden biases
employed
then apply to extreme learning machine. Two real public traffic monitoring datasets from Internet service providers
to evaluate our design. The achieved outcomes demonstrate that our proposed solution works effectively with satisfied performance were employed
to
in evaluate our with
comparison design. The achieved
existing outcomes
models for demonstrate
distributed that our proposed
system workload solution works effectively with satisfied performance
prediction.
in comparison with existing models for distributed system workload prediction.
c 2020

© 2020 The Authors. Published
The Authors. Published by
by Elsevier
Elsevier B.V.
B.V.
c 2020an

This The Authors. Published by Elsevier B.V.
This is
is an open
open access
access article
article under
under the
the CC
CC BY-NC-ND
BY-NC-ND license
license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
This is an open
Peer-review
Peer-review access
under article under
responsibility
responsibility theConference
of the
of the CC BY-NC-ND
Conference license
Program
Program (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Chairs.
Chairs.
Peer-review under responsibility of the Conference Program Chairs.
Keywords: meta-heuristics; opposition-based tug of war optimization; extreme learning machine; internet traffic; workload prediction
Keywords: meta-heuristics; opposition-based tug of war optimization; extreme learning machine; internet traffic; workload prediction

1. Introduction
1. Introduction
Time series data is a sequence of related data points gathered over time. The goal of time series data prediction
Timeisseries
process dataaisforecast
to make a sequence of points
of data relatedindata
the points gathered
near future over
based on time. The goaldata.
the historical of time
Thisseries datais prediction
process useful for
process is to make a forecast of data points in the near future based on the historical data. This
many commercial and industrial applications such as economy forecast, sale improvement, budgetary analysis, process is useful for
stock
many commercial and industrial applications such as economy forecast, sale improvement, budgetary
trading, yield projections, process and quality control of systems, inventory studies, workload projections, census analysis, stock
trading,
analysis yield projections,
and many processthere
more. Because and are
quality
manycontrol of systems,
methods, which areinventory studies,
used to model andworkload projections,
predict time census
series data, two
analysis and many more. Because there are many methods, which are used to model and predict time series
criteria that are often taken into consideration when applying a forecast technique, including accuracy and effective- data, two
criteria that are often taken into consideration when applying a forecast technique, including accuracy and effective-

∗ Corresponding author. Tel.: +84-24-3869-6124 ; fax: +84-24-3869-2906

∗ Corresponding
E-mail address:author. Tel.: +84-24-3869-6124 ; fax: +84-24-3869-2906
[email protected]
E-mail address: [email protected]
1877-0509  c 2020 The Authors. Published by Elsevier B.V.
1877-0509
This c 2020

is an open Thearticle
access Authors. Published
under by Elsevier B.V.
the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
1877-0509
This
Peer-review ©
is an open2020 Thearticle
access
under Authors. ofPublished
under
responsibility by Elsevier
the Conference
the CC BY-NC-ND B.V.Chairs.
license
Program (http://creativecommons.org/licenses/by-nc-nd/4.0/)
This is an open access article under the CC BY-NC-ND
Peer-review under responsibility of the Conference Program Chairs.license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the Conference Program Chairs.
10.1016/j.procs.2020.03.063
 Thieu Nguyen et al. / Procedia Computer Science 170 (2020) 362–369 363
2 T. Nguyen et al. / Procedia Computer Science 00 (2019) 000–000

ness [1, 23]. While the precision can be measured by many standard metrics according to each prediction method,
the effectiveness is difficult to quantify by parameters. For instance, in addition to the accuracy, simpleness, learning
speed, the computation cost are also important factors that affects the effectiveness of time series prediction models.
In this paper, we focus on developing a novel time series prediction technique, which can tackle the non-linear
models and achieve the same level or even better accuracy, stability, time, and interpretability as compared with other
modern methods. The prediction model thus can be used to many applications, especially in this work we use for fore-
casting workloads of distributed systems. In this direction, we contribute a new version for extreme learning machine
(ELM) using our proposed opposition-based tug of war optimization (OTWO) algorithm to select input weights, and
using MoorePenrose generalized inverse to calculate output weights. The new gained prediction technique is called
OTWO-ELM. The reasons for ELM choice is that this model has several advantages such as simple, interpretable,
fast learning, and can bring good quality forecast results. In our model, the improvement of OTWO mainly focuses on
decreasing the norm of output weights of ELM and constraining input weights and hidden biases within a reasonable
range to enhance the convergence performance of that neuron network. Hence, it improves the prediction accuracy
for our model. OTWO-ELM was experimented and evaluated with two real datasets covering the Internet traffic data
from European Union (EU) and United Kingdom (UK) cities provided by Internet service providers. We also com-
pared the obtained performance of OTWO-ELM with several other well-known prediction techniques. The gained
results showed that our proposals bring positive effectiveness in predicting workloads for distributed systems.
The structure of this paper is as follows. In Section 2, we classify and analyze existing studies to highlight our
contributions. Section 3 describes OTWO-ELM design. In Section 4, we present experiments and evaluations for the
proposed model to prove its effectiveness. Section 5 concludes and defines our future work directions.

2. Related Work

Artificial neural networks are an effective universal approximator widely used for machine learning regression
and classification [29]. However, artificial neural networks often use gradient-based learning algorithms like back-
propagation, which causes many drawbacks [2], for example, easily trapped in local minimum, time-consuming com-
putation due to improper learning steps. In addition, the performance of artificial neural networks often is not very
stable [32]. Meanwhile, deep neural networks have more hidden layers and bring good performance [17]. The most
well-known deep neural networks type for time series forecasting is recurrent neural network [7] as well as its variants
like long short term memory [14], and gated recurrent unit [3]. However, there are still two main concerns toward
deep neural network usage: 1) requirement of a larger amount of data to train [33] in order to gain more accuracy and
2) the computational cost due to the model complexity [26].
In order to overcome drawbacks and concerns, we consider ELM [10] in this work. It randomly chooses the
input weights and hidden biases, determines the output weights of single-layer feed-forward networks through the
simple generalized inverse operation of the output matrix of hidden layer. ELM not only learns much faster with
better generalization performance but also avoids many difficulties faced by gradient-based learning methods such
as stopping criteria, learning rate, learning epochs, and local minimum. The interpretable of ELM has been proved
in term of both universal approximation and classification capabilities in [8, 9]. However, ELM still tends to require
more hidden neurons than traditional gradient-based learning algorithms. This also leads to ill-condition problem
due to randomly selecting input weights and hidden biases. In [34], the authors proposed an ELM version using the
differential evolutionary algorithm to select input weights, and using MoorePenrose generalized inverse to analytically
determine output weights. These improvements can bring good performance and make a compact ELM network. In
this direction, several meta-heuristic algorithms have been proposed to combine with ELM such as particle swarm
optimization (PSO) [30], genetic algorithm (GA) [31] to improve ELM performance. However, PSO weaknesses are
pointed out [16, 25] as easily to be trapped in the local optima when solving complex multi-modal problems. For
GA, many studies such as [20, 24] identified its issues. Concretely, with the high epistatic targeting functions, the
technique performance degradation is often large. It is suitable to mention hyper-heuristic strategy for hybridization
of nature inspired algorithms like [6], which is designed to efficiently solve computational search problems.
Recently, a novel meta-heuristic called tug-of-war optimization (TWO) based on physics laws in the tug of war
game in competitions was proposed in [11]. The authors have proved their effectiveness in several aspect such as
solving the structural damage identification problems [13], applying to the optimal design of castellated beams [12].
364 Thieu Nguyen et al. / Procedia Computer Science 170 (2020) 362–369
T. Nguyen et al. / Procedia Computer Science 00 (2019) 000–000 3

Even with high performance in solving some problems, like other meta-heuristic algorithms, the traditional TWO can
get stuck in local optimal point, and therefore the convergence becomes slow and time-consuming. These problems
occur because some solutions update their positions based on the best solution. However, the optimal solution might
be in the opposite direction to the current solution. Considering such drawbacks, we propose an improvement of TWO
by searching global solution in both directions (forward and opposite). This proposal thus is called opposition-based
tug of war optimization (OTWO for short). Based on our knowledge, there is no research to enhance TWO by using
opposition-based learning, and no study improves the ability of ELM by OTWO. Our main contributions include:

1. Proposing an improvement of tug of war optimization based on opposition-based learning called OTWO;
2. Proposing an improvement of ELM which using opposition-based tug of war optimization to find good input
weights and hidden biases (called OTWO-ELM);
3. Experimenting and evaluating several models including well-known multilayer neural network (MLNN), tradi-
tional ELM, GA-ELM, PSO-ELM, TWO-ELM and OTWO-ELM for the system workload prediction problem
(in this study, Internet traffic is used as test datasets).

3. Designing Opposition-based Tug of War Optimization Extreme Learning Machine

3.1. Tug of War Optimization

Tug of War Optimization (TWO) is a multi-agent meta-heuristic technique, which simulates each solution Xi = xi, j
(Eq. 1) as a team, which participates in a series of tug of war matches. The team’s weight is defined based on the
quality of that corresponding solutions (Eq. 1), and the amount of pulling force, which a team can deploy on the rope,
which is assumed to be proportional to its weight (Eq. 2). Naturally, a lighter weight team will be pulled towards
a stronger weight team. The principle helps form the converging operator based on Newtonian (mechanics) laws in
TWO algorithm (Eq. 2 to 4). The convergence operator will improve the quality of each solution based on maintaining
a reasonable balance between exploration/exploitation phase. The details of TWO is presented in [11].

 
f iti − f itworst
xi0j = x j,min + rand(x j,max − x j,min ) j = 1, 2, ..., n Wi = 0.9 + 0.1 i = 1, 2, .., N (1)
f itbest − f itworst
F kp,i j = max{Wi µ s , W j µ s } k
Fr,i k k
j = F p,i j − Wi µk (2)
k
Fr,i j
gkij = X kj − Xik akij = gkij (3)
Wik µk
N

1 k 2
∆Xikj = a ∆t + αk β(Xmax − Xmin )  randn(1, n) ∆Xik = ∆Xikj (4)
2 ij j=1
N
 1
Xik+1 = Xik + ∆Xik = Xik + ∆Xikj xik+1 k
j = GB j + randn(GB j − xi j ) (5)
j=1
k

where xi0j is the initial values of jth variable of the ith candidate solution; x j,max , x j,min are the maximum and minimum
values of jth variable respectively; rand is random number from a uniform distribution in the interval [0, 1]; n is the
number of optimization variables; f iti is the fitness value of ith solution; f itworst , f itbest are the worst and the best
fitness values of teams; µ s , µk are the static coefficient and kinematic coefficient of friction; Wµ s is the pulling force
is assumed equal to its static friction force; F kp,i j is the pulling force between teami and team j (Newton’s third law);
k th k
Fr,i j is the resultant force of heavier team j on teami in k iteration; gi j is the gravitational acceleration constant in
k k
Newton’s second law; Xi , X j are the position vector of solution i and j in k iteration; ∆Xikj is the displacement of teami
after competing with team j ;αk is to gradually decrease the random portion of the team’s movement; αk ∈ [0.9, 0.99];
β is a scaling factor which can be chosen from the interval (0, 1].  is the element wise element multiplication; Xik+1
is the new position of teami after a complete round; GB j is the best solution so far; randn is a random number drawn
form a standard normal distribution.
 Thieu Nguyen et al. / Procedia Computer Science 170 (2020) 362–369 365
4 T. Nguyen et al. / Procedia Computer Science 00 (2019) 000–000

3.2. Opposition-based Tug of War Optimization

Opposition-based Learning (OBL) is used to improve the convergence of meta-heuristic methods to find the global
solution of the optimization problem [28]. In general, OBL indicates that for finding the unknown optimal solution,
searching both a random direction and its opposite simultaneously gives a higher chance to find the promising regions
and to enhance the algorithm performance [15]. In the way, our improvement for TWO covers two stages above.
Firstly, the OBL is employed in the initialization of teams to enhance the rate of convergence and avoid stuck in local
optima through searching solutions in the whole domain. Secondly, in the updating stage of the population solution,
the OBL is used to check if the update in the opposite direction is better than the current ones. These two stages are
explained in detail as follows.
Initialization Stage: In general, meta-heuristic algorithms usually begin with initializing a random population,
which is an important starting point. If the diversity of the initial population is poor or all solutions are in the local
optima region, the algorithms may be led to early convergence and converge at local optima. To ensure the appropriate
diverse for the population, we propose the following schema: 1) Randomizing teams (population) S of size N/2 based
on Eq. 1. 2) Generating opposite teams S using OBL to create the opposite solution in S using Eq. 6. 3) With two
populations S and S , we have N randomly solutions.
Update Stage: In this stage, once the teams of a league compete against each other for a complete round, the
league should be updated. This is done by comparing the new candidate solutions (the new positions of the teams) to
the current teams of the league. If the new candidate solution i is better than the old one in terms of fitness value, the
old one is replaced by the new solution. If not, we generate the opposite solution of the old solution using Eq. 7, then
comparing the fitness between the old solution and the opposite of the old solution. After this stage, which solution
has better fitness value will be kept.
x i j = u j + l j − xi j , i = 1, 2, .., N and j = 1, 2, ..., n (6)
xi j = u j + l j − GBi j + rand(GBi j − xi j ) (7)

where xi j is the ith point of the jth solution of S ; xi j is the opposite solution of the xi j solution of S ; u j and l j are the
upper and lower bound of the jth variable, respectively; GBi j is the global best of the current iteration.
The final development of OTWO is illustrated in the Algorithm 1.

3.3. Opposition-based Tug of War Optimization Extreme Learning Machine (OTWO-ELM)

Our proposed prediction model includes two main algorithms, which are ELM and OTWO as presented in Fig. 1.
As mentioned in related work section, ELM was proposed to overcome the drawbacks of gradient-based methods
in single-hidden layer feed-forward neural networks by implementing two phases. 1) Phase 1: calculating weights
between input and hidden layer by choosing randomly; 2) Phase 2: calculating weights between hidden and output
layer by simple generalized inverse operation of the hidden layer output matrix.
The strength of ELM is speed because it requires little time to learn the relation between input and output by random
process. However, this is a trade-off between the speed and generalization performance. Non-optimal input weights
may be randomly chosen, and this causes bad performance. In order to tackle the problem, in this paper, OTWO
is used to replace the random process in the Phase 1 (Fig. 1). There are two important operations in Fig. 1, which
are encoding and decoding. The encoding operation encodes weights and biases of the input and hidden layer into
a solution/team (real-value vector). Otherwise, the decoding operation decodes the solution into weights and biases
of input and hidden layer. In order to avoid long training time problem, a termination criterion with completing a
maximum epochs number (training cycles) or achieving error goal is designed. Operations of OTWO in OTWO-ELM
model are introduced in Algorithm 1.

4. Experiment and Evaluation

There are two aspects that we consider in our experiments to evaluate the proposals, namely prediction accuracy and
stability of models. In this way, tested prediction models consist of well-known multi-layer neuron network (MLNN),
traditional ELM, several variants of ELM cover GA-ELM, PSO-ELM, TWO-ELM, and our proposed OTWO-ELM.
The related materials and the implementation of our experiment can be found in [19, 18].
366 Thieu Nguyen et al. / Procedia Computer Science 170 (2020) 362–369
T. Nguyen et al. / Procedia Computer Science 00 (2019) 000–000 5

Algorithm 1 Opposition-based Tug of War Optimization

Input: The population size p s , the maximum number of generations gmax
Output: The best team
1: Initialize population following the strategy ”Initialization stage” in 3.2
2: Calculate the weights W of the teams based on Eq. 1 (part 2)
3: Find the global best GB solution
4: g←1
5: while g ≤ gmax do
6: for each teami do
7: for each team j do
8: if Wi < W j then
9: Move teami towards team j using Eq. 4 (part 2)
10: Find the total displacement of teami based on Eq. 4
11: Determine the final position of teami using Eq. 5 (part 1)
12: Amend solution when its outside of the search space using Eq. 5
13: for each teami do
14: if teamnew
i has better fitness than teamold
i then
15: Replace teamold
i by teami
new

16: else
17: Find the opposite position of teamold
i based on Eq. 7
18: Keep teamold
i or replace by the opposite one based on fitness
19: Update team weights
20: Find the current best solution
21: Keep the global best or replace by the current best solution based on fitness
22: Return GB

bias weights
α 01 α 0q α 11 α 1q α p1 α pq
No

Team / Initialize Opposition-based Tug of Are optimization /

Solution Population War Optimization termination criteria
met?
Encoder

Output Training Data Yes

Input Hidden

bias β01
bias β11 xO1 yt
Calculate output
Best Team
α 01
weights by
α 11 xH1 β21 β
q1 Moore-Penrose
α p1 yt+1 inverse matrices Decoder
yt-p
xI1 xO2
xH2
Input

Input weights
β1w β0w
yt-1 α 0q β2w
xIp α 1q xOw yt+w-1
βqw Trained Model
α pq xHq

Single-Hidden Layer Feedforward Neural Network Structure

Fig. 1: OTWO-ELM training process

Datasets: In this work, we used two datasets [5]. The Internet traffic (in Megabyte) from data center in European
(EU) cities and Internet traffic (in Bytes) from United Kingdom (UK) academic network backbone. Both datasets were
collected at five-minute intervals. Both datasets are preprocessed into the form of time-series through normalization
 Thieu Nguyen et al. / Procedia Computer Science 170 (2020) 362–369 367
6 T. Nguyen et al. / Procedia Computer Science 00 (2019) 000–000

Fig. 2: Visualize the results of OWTO-ELM forecasting for EU (left) and UK (right) dataset

and transformation [22] before being used. In our tested models, we use Root Mean Square Error (RMSE) to
evaluate the error between output of traditional network and true value, and the fitness function in meta-heuristic
algorithms. For more objective, we also compare the gained results using Mean Absolute Percentage Error (MAPE)
and Coefficient of Determination (R2 ).

Setting Parameters: The train:test ratio is set to 0.8:0.2 for all datasets

1. Default model settings: The number of layers of MLNN, ELM and versions of ELMs are set to three layers (1
input, 1 hidden and 1 output); The input size p = km, where k is the sliding window size; m is dataset metric
number (univariate m = 1, multivariate m > 1); The size of the hidden layer is set to 20 neurons for all models.
This setting is chosen small enough to avoid over-fitting and big enough to well-learning the mapping function
from inputs to outputs. The output size w is 1. Activation function is set to Exponential Linear Unit (ELU) [4].
2. Meta-heuristics algorithms: For all meta-heuristics algorithms (GA, PSO, TWO, OTWO), we set the population
at 20 and the number of generation is 100; GA setting is based on the work [21] with crossover ratio pc is 0.95
and mutation ratio pm is 0.025; PSO seting is based on [27]: inertia factor w is set as linearly reducing with
iteration from 0.9 to 0.4, cognitive learning rate c1 = c2 = 1.2; TWO and OTWO have similar parameters, which
are set based on the original paper [? ] for Eq. 4: α = 0.99, scaling factor β = 0.1, and ∆t = 1.
3. The sliding window size is set to 2 and 5. The size is 2 that means two data points at time step (t − 2) and (t − 1),
i.e., (xt−2 , xt−1 ) are used to predict data point at time step t. The sliding window at 5 is set in a similar way.

4.1. Prediction Accuracy

Table 1 (left side) shows obtained results when running models with EU and UK datasets with different measure-
ment methods. Meantime, Fig. 2 illustrates the prediction outcomes against the real values of testing models.
With EU dataset. When sliding window k = 2, our proposed model achieves the two best results as compared
with others with RMSE and R2 measurements. The two best results are 15.363 and 0.9958 for RMSE and R2 , re-
spectively. For MAPE, OTWO-ELM achieves the second-best result that is only after TWO-ELM. When k = 5,
OTWO-ELM gives the best results in MAPE and R2 metrics with values of 2.853 and 0.9959, respectively. With
RMSE, OTWO-ELM gives the second-best result after PSO-ELM.Hence, it can be made an observation as follows:
with all measurement metrics, the obtained results when running MLNN are bad. ELM is slightly better than MLNN.
Thus, the values of that metrics are significantly better for combination models among meta-heuristic algorithms and
ELM. The outcomes from traditional TWO are quite similar when compared with GA and PSO techniques. However,
when combining Opposition-based learning into TWO, the novel OTWO bring the best results as compared with the
others in most experiments.
With UK dataset. For OTWO-ELM model, the obtained results show even higher performance in comparison
with EU dataset tests. While k = 2, OTWO-ELM makes the best results for all metrics including RMSE, MAPE, and
368 Thieu Nguyen et al. / Procedia Computer Science 170 (2020) 362–369
T. Nguyen et al. / Procedia Computer Science 00 (2019) 000–000 7

Table 1: Prediction accuracy and the stability between our proposed OTWO-ELM and other models on different datasets

Prediction Accuracy Stability (based on MAPE)

Data Model k=2 k=5 Statistics (15 runtimes)
RMSE MAPE R2 RMSE MAPE R2 min max mean std cv
MLNN 16.716 3.0125 0.9915 15.857 2.9826 0.9927 2.869 3.157 2.965 0.108 0.036
ELM 16.292 2.9834 0.9948 15.744 2.9255 0.9951 2.882 3.189 2.919 0.077 0.027
GA-ELM 15.378 2.9203 0.9956 14.914 2.8591 0.9958 2.836 3.054 2.882 0.064 0.022
EU
PSO-ELM 15.394 2.9255 0.9955 14.834 2.8856 0.9959 2.837 2.935 2.863 0.029 0.010
TWO-ELM 15.400 2.9171 0.9955 14.885 2.8645 0.9958 2.843 2.883 2.859 0.011 0.004
OTWO-ELM 15.363 2.9249 0.9958 14.874 2.853 0.9959 2.834 2.873 2.854 0.011 0.004
MLNN 11.183 1.5687 0.9958 11.925 1.5945 0.9966 1.395 1.737 1.487 0.108 0.073
ELM 10.511 1.449 10.190 0.9974 1.4119 0.9975 1.395 1.652 1.460 0.069 0.047
GA-ELM 10.716 1.4157 0.9975 10.364 1.3399 0.9977 1.330 1.398 1.353 0.025 0.01
UK
PSO-ELM 10.696 1.4126 0.9975 10.112 1.3701 0.9976 1.327 1.402 1.347 0.026 0.019
TWO-ELM 10.714 1.4170 0.9975 10.410 1.3781 0.9976 1.329 1.385 1.340 0.017 0.013
OTWO-ELM 10.349 1.3677 0.9976 10.115 1.3383 0.9977 1.324 1.361 1.337 0.009 0.007

R2 . Otherwise, when k = 5, OTWO-ELM brings the best results with MAPE, R2 measurements, and gives the second-
best result just after PSO-ELM with RMSE. Note that the difference between RMSE’s best result and our proposed
OTWO-ELM is very small but there is a huge difference as compared with others. There is also an observation here:
the experimental results got with this data is like to EU data. Hybrid models of TWO-ELM, GA-ELM, PSO-ELM
produce quite similar results. However, the proposed model OTWO-ELM has better outcomes as compared with other
models.

4.2. Stability

Table 1 shows Min, Max, Mean, Standard deviation (Stdev) and Coefficient of Variation (CV) of MAPE values
after 15 runtimes with both test data. There are some remarks from the experiments. With UK dataset, the values of
Min and Max of OTWO-ELM are the smallest. It means that in the best of worst cases, the OTWO-ELM offers the best
results as compared to the others. The Mean values of OTWO-ELM are also the smallest value. The minimum Stdev
and CV values also indicate that OTWO-ELM model is the most stable. With EU dataset, it can be seen that OTWO-
ELM gives the best results in all 5 values Min, Max, Mean, Stdev. This again proves that our proposed OTWO-ELM
outperform in both performance and stability perspectives.

5. Conclusion and Future Work

In this paper, the novel prediction model (OTWO-ELM) for time series data is proposed and validated with experi-
ments. The main contributions of the work include 1) developing a novel version of tug of war optimization based on
opposition-based learning; 2) proposing a new prediction model for time series data using a novel variants of extreme
learning machine; 3) carrying out experiments and evaluations for OTWO-ELM with real datasets of Internet traffic.
The advantages of the proposed model is not only the simplicity, but also its good accuracy, stability, minimal com-
putation cost and interpretable as compared with others. In the future, the approach will be enhanced with more core
methods to train different neural network types.

Acknowledgements

The research is funded by the Hanoi University of Science and Technology under project number T2018-PC-017.
It is also supported by the projects VEGA 2/0125/20 and APVV-17-0619.
 Thieu Nguyen et al. / Procedia Computer Science 170 (2020) 362–369 369
8 T. Nguyen et al. / Procedia Computer Science 00 (2019) 000–000

References

[1] Ahmad, A., Javaid, N., Guizani, M., Alrajeh, N., Khan, Z.A., 2016. An accurate and fast converging short-term load forecasting model for
industrial applications in a smart grid. IEEE Transactions on Industrial Informatics 13, 2587–2596.
[2] Aljarah, I., Faris, H., Mirjalili, S., 2018. Optimizing connection weights in neural networks using the whale optimization algorithm. Soft
Computing 22, 1–15.
[3] Chung, J., Gulcehre, C., Cho, K., Bengio, Y., 2014. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv
preprint arXiv:1412.3555 .
[4] Clevert, D.A., Unterthiner, T., Hochreiter, S., 2015. Fast and accurate deep network learning by exponential linear units (elus). arXiv preprint
arXiv:1511.07289 .
[5] Cortez, P., Rio, M., Rocha, M., Sousa, P., 2012. Multi-scale internet traffic forecasting using neural networks and time series methods. Expert
Systems 29, 143–155.
[6] Damaševičius, R., Woźniak, M., 2017. State flipping based hyper-heuristic for hybridization of nature inspired algorithms, in: International
conference on artificial intelligence and soft computing, Springer. pp. 337–346.
[7] Gal, Y., Ghahramani, Z., 2016. A theoretically grounded application of dropout in recurrent neural networks, in: Advances in neural information
processing systems, pp. 1019–1027.
[8] Huang, G.B., Chen, L., Siew, C.K., et al., 2006a. Universal approximation using incremental constructive feedforward networks with random
hidden nodes. IEEE Trans. Neural Networks 17, 879–892.
[9] Huang, G.B., Zhou, H., Ding, X., Zhang, R., 2011. Extreme learning machine for regression and multiclass classification. IEEE Transactions
on Systems, Man, and Cybernetics, Part B (Cybernetics) 42, 513–529.
[10] Huang, G.B., Zhu, Q.Y., Siew, C.K., 2006b. Extreme learning machine: theory and applications. Neurocomputing 70, 489–501.
[11] Kaveh, A., 2017. Tug of war optimization, in: Advances in Metaheuristic Algorithms for Optimal Design of Structures. Springer, pp. 451–487.
[12] Kaveh, A., Shokohi, F., 2016. Optimum design of laterally-supported castellated beams using tug of war optimization algorithm. Structural
Engineering and Mechanics 58, 2016.
[13] Kaveh, A., Zolghadr, A., 2017. Guided modal strain energy-based approach for structural damage identification using tug-of-war optimization
algorithm. Journal of Computing in Civil Engineering 31, 04017016.
[14] Ma, X., Tao, Z., Wang, Y., Yu, H., Wang, Y., 2015. Long short-term memory neural network for traffic speed prediction using remote microwave
sensor data. Transportation Research Part C: Emerging Technologies 54, 187–197.
[15] Mahdavi, S., Rahnamayan, S., Deb, K., 2018. Opposition based learning: A literature review. Swarm and evolutionary computation 39, 1–23.
[16] Mirjalili, S., Lewis, A., 2016. The whale optimization algorithm. Advances in engineering software 95, 51–67.
[17] Nguyen, A., Yosinski, J., Clune, J., 2015. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images, in:
Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 427–436.
[18] Nguyen, T., 2020a. Opfunu: Python library for optimization functions in numpy. doi:10.5281/zenodo.3620961.
[19] Nguyen, T., 2020b. Otwo-elm: Opposition-based tug of war optimization - extreme learning machine. doi:10.5281/zenodo.3626115.
[20] Nguyen, T., Nguyen, B.M., Nguyen, G., 2019a. Building resource auto-scaler with functional-link neural network and adaptive bacterial
foraging optimization, in: International Conference on Theory and Applications of Models of Computation, Springer. pp. 501–517.
[21] Nguyen, T., Nguyen, T., Nguyen, B.M., Nguyen, G., 2019b. Efficient time-series forecasting using neural network and opposition-based coral
reefs optimization. International Journal of Computational Intelligence Systems 12, 1144–1161. doi:10.2991/ijcis.d.190930.003.
[22] Nguyen, T., Tran, N., Nguyen, B.M., Nguyen, G., 2018. A resource usage prediction system using functional-link and genetic algorithm neural
network for multivariate cloud metrics, in: 11th Conference on Service-Oriented Computing and Applications (SOCA), IEEE. pp. 49–56.
[23] Oliveira, T.P., Barbar, J.S., Soares, A.S., 2016. Computer network traffic prediction: a comparison between traditional and deep learning neural
networks. International Journal of Big Data Intelligence 3, 28–37.
[24] Połap, D., Woźniak, M., et al., 2019. Bio-inspired voice evaluation mechanism. Applied Soft Computing 80, 342–357.
[25] Połap, D., et al., 2017. Polar bear optimization algorithm: Meta-heuristic with fast population movement and dynamic birth and death mecha-
nism. Symmetry 9, 203.
[26] Qin, Y., Song, D., Chen, H., Cheng, W., Jiang, G., Cottrell, G.W., 2017. A dual-stage attention-based recurrent neural network for time series
prediction, in: IJCAI-17, pp. 2627–2633. doi:10.24963/ijcai.2017/366.
[27] Salih, S.Q., Alsewari, A.A., 2019. A new algorithm for normal and large-scale optimization problems: Nomadic people optimizer. Neural
Computing and Applications , 1–28.
[28] Tizhoosh, H.R., 2005. Opposition-based learning: a new scheme for machine intelligence, IEEE. pp. 695–701.
[29] Walczak, S., 2019. Artificial neural networks, in: Advanced Methodologies and Technologies in Artificial Intelligence, Computer Simulation,
and Human-Computer Interaction. IGI Global, pp. 40–53.
[30] Xu, Y., Shu, Y., 2006. Evolutionary extreme learning machine–based on particle swarm optimization, in: International Symposium on Neural
Networks, Springer. pp. 644–652.
[31] Yang, H., Yi, J., Zhao, J., Dong, Z., 2013. Extreme learning machine based genetic algorithm and its application in power system economic
dispatch. Neurocomputing 102, 154–162.
[32] Zhang, H., Wang, Z., Liu, D., 2014. A comprehensive review of stability analysis of continuous-time recurrent neural networks. IEEE
Transactions on Neural Networks and Learning Systems 25, 1229–1262.
[33] Zhou, P., Shi, W., Tian, J., Qi, Z., Li, B., Hao, H., Xu, B., 2016. Attention-based bidirectional long short-term memory networks for relation
classification, in: Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2), pp. 207–212.
[34] Zhu, Q.Y., Qin, A.K., Suganthan, P.N., Huang, G.B., 2005. Evolutionary extreme learning machine. Pattern recognition 38, 1759–1763.