You are on page 1of 6

Recent Trends in Nature Inspired Computation

with Applications to Deep Learning


Vandana Bharti Bhaskar Biswas Kaushal Kumar Shukla
Computer Science and Engineering Computer Science and Engineering Computer Science and Engineering
Indian Institute of Technology (BHU) Indian Institute of Technology (BHU) Indian Institute of Technology (BHU)
Varanasi, Uttar Pradesh, India Varanasi, Uttar Pradesh, India Varanasi, Uttar Pradesh, India
vandanabharti.rs.cse17@iitbhu.ac.in bhaskar.cse@iitbhu.ac.in kkshukla.cse@iitbhu.ac.in

Abstract—Nature-inspired computations are commonly rec- Bat algorithm, Cuckoo search algorithm (CS), Artificial bee
ognized optimization techniques that provide optimal solutions colony, Harmony search (HS), Ant colony optimization (ACO)
to a wide spectrum of computational problems. This paper and Invasive weed optimization algorithm are just a few to
presents a brief overview of current topics in the field of nature-
inspired computation along with their most recent applications mention. At this stage, there is a need to discuss in holistic
in deep learning to identify open challenges concerning the manner about principles, ideas, applications, problems and
most relevant areas. In addition, we highlight some recent solutions linked to the use of NIA.
hybridization methods of nature-inspired computation used to Recently, Deep Learning (DL) has emerge as a hot topic,
optimize the hyper-parameters and architectures of a deep presenting impressive results with NIA. DL, a buzzword for
learning framework. Future research as well as prospective deep
learning issues are also presented. today’s era of artificial intelligence (AI) researchers, involves
Index Terms—nature-inspired computation, neuroevolution, machine learning (ML) methods that emphasize high-level
dynamic optimization, evolutionary algorithms, deep learning data representation learning rather than task-specific learning
[5]. DL has been extensively used due to advances in hardware
I. INTRODUCTION and the development of computational fast and efficient
There is a wide range of high-performance optimization optimization algorithms [6]. In addition, DL can be used to
methods available for problems formulated as convex opti- evaluate and mine exciting Big Data models [7]. Like several
mization [1]. However, many problems arising in important other fields of problem, DL also consist of an optimization
application domains pose non-convex problems. In particular, problem within itself which may be single or multi-objective,
these problems are often noisy, stochastic and have limitations constrained or unconstrained and convex or non-convex.
that change dynamically. This is where nature inspired The application diversity of NIC algorithms is enormous and
computation (NIC) methods are often applied as they have the literature is growing quickly. Therefore, this paper only
been shown to be highly successful when dealing with such tries to present a state-of-the-art of NIC latest developments
difficult problem domains. NIC is the generalization of bio- with its most recent application for DL to discover the optimal
inspired computing and evolutionary computing where, in each structure of DL. Furthermore, we also present the work done
case, algorithms are derived from the natural phenomenon to optimize a DL model’s parameters and hyperparameters for
of the surrounding environment, nature or habitats. These a specified search space within a sensible and reasonable span
computing methods have led to the formation of algorithms of time. In addition to exploring the recent work on applying
called Nature Inspired Algorithms (NIA) [2]. these optimization techniques to DL, this paper also discusses
NIA have achieved enormous popularity in latest years some significant future research directions in order to enhance
in tackling problems belong to NP hard (NPH) and NP synergy between NIC and DL.
complete (NPC) complexity classes and solving hard opti- The paper is organized as follows: Section-2 presents new
mization tasks whose real solution is unknown. The objective trends in NIC. The most recent applications of NIC in DL
of developing such algorithms is to optimize engineering are discussed in Section-3 followed by conclusion and future
problems. Engineering problems are becoming progressively direction in Section-4.
complex and difficult to optimize as the world moves towards
industrialisation. In order to deal with these scenario, NIA are II. W HATS N EW IN NATURE -I NSPIRED C OMPUTATION
crafted and tested on numerical benchmark functions [3] for NIC sought to develop novel computing methods by
single and multi-objective optimization, and designed to solve monitoring how natural phenomena act to solve complex
NPH problems with high-dimensions. problems in different environmental circumstances.The pri-
A lot of promising NIAs have appeared over the last mary goal of NIC is to create computational models and
two decades [4]. In this respect, Grey wolf optimiza- efficient algorithms based on natural intelligence discovered
tion (GWO), Whale Optimization Algorithm (WOA), Harris in chemical, physical and biological systems. This has led to
Hawks Optimizer (HHO), Particle swarm optimization (PSO), ground-breaking studies that has created new branches such as

978-1-7281-2791-0/20/$31.00 2020
c IEEE 294
Fig. 1: Graphical representation of the areas within nature inspired computation and some of their recent applications.
neuro-evolution, swarm intelligence, evolutionary computing effective optimizers inspired by natural phenomenon, animal
and artificial immune systems. NIC techniques are applied behavior and physical science [16], [17]. Apart from these,
in the areas of physics, biology, engineering, economics, and the new popular research area emerges as an introduction of
even management. Chaos theory in NIA. Chaos theory has random, dynamic, non-
There is a wide spectrum of domains such as transportation, repetitive and ergodic properties. Due to its dynamic property
portfolio optimization, supply chain management, agriculture, it ensures the different solutions given by algorithms even
healthcare, remote sensing, industrial problems, complex on the complex multi-modal landscape. Since, most of the
networks like social networks, wireless sensor network, and NIA belongs to stochastic algorithms in which randomness
many more, attract the research community to solve these is achieved by using probabilistic theory such as Gaussian
as an optimization problem. Graphical representation of the distribution or uniform distribution now a days to accelerate
areas within nature inspired computation are depicted in Fig.1. the convergence and to enhance the diversity of the NIA
After observing recent research and application, we have chaos theory is used. Due to its ergodicity its performs
divided the work into two subclasses, one of which involves searches at higher speed. This novel approach has been
theoretical studies such as convergence analysis, stability widely utilized in various NIA and its applications [18].
analysis, and multidisciplinary studies with NIA and other Another challenging area is convergence analysis which is
belongs to real-world applications of NIA embedded ML and still unexplored. Only few works are there that tried to solve
DL models. Multidisciplinary studies integrate concepts of this by using markov chain model [19]. Hybrid model of
distinct domains with these nature-inspired techniques and existed NIA to enhance the performance of optimizer and
resulted in new better performing optimizer such as CHAOS ML embedded with NIA are few recent development which
theory from mathematics is integrated with several swarm are used either for feature selection or model parameter
based algorithms [8]–[10] superposition theorem, big bang optimization and solving several real world applications [20],
theory and black hole from physics give rise to new optimizer [21]. Self adaptive NIA [22], cellular automata based NIA
[11], [12]. Recent trends in this domain show that research [23], [24], surrogate assisted NIA [25] and new innovation in
communities are attracted towards multidisciplinary studies multiobjective optimization and many objective optimization
and their integration with model to improve the quality are the new attraction for solving versatile application [26],
of solution and try to reduce the exhaustive computation. [27]. There is a continuous innovation in technology which
Moreover, several researches demonstrate that researchers are results more complex problems and need better optimizer
now solving the real world and challenging applications by which leads to design new complex benchmarks such as high
treating them as an optimization problem that can be a single- scale benchmarks for validating optimizer [28]. From last few
objective, many-objective, dynamic objective, hierarchical in months it has been observed that DL which is highly popular in
nature and many more [13]–[15]. almost every domain of application also integrates with nature
Furthermore, latest work demonstrates that research com- inspired optimizers and here it comes new complex,interesting
munities are continuously striving to develop promising and and challenging area to be explored more.

10th International Conference on Cloud Computing, Data Science & Engineering (Confluence) 295
Fig. 2: Applications of nature inspired computations in Deep Learning.

III. R ECENT A PPLICATIONS OF NIC IN D EEP L EARNING Search, CS, GWO, PSO, and applied on cervical cancer that
NIC is a fascinating new area along with DL that promises achieved accuracy above than 96% in [29]. In another research
real-world applications with things like threat detection, mo- work Li et al. [30] proposed competitive random search based
tion tracking, robotics, medical imaging, aerodynamic design on GA and also proposed Evolutionary Attention based LSTM
optimization, source localization and many more. Recently, (EA-LSTM) for multivariate time series forecasting. Results
neuro-evolution popularized by Stanley and Miikkulainen have shown that the use of competitive random search to
with DL gained extensive popularity, refers to the artificial train EA-LSTM not only enables long-term dependence on
intelligence paradigm that concerns the design of ANNs, time series prediction, but can also effectively use local data
including their topology, optimum weights and learning rules, in a sampling window based on different scales of attention
using simulated evolution. distribution. The proposed model also delivered robust and
DL is a branch of machine learning that imitates the competitive performance as compare to other baseline method.
thought process of human brain to learn high level abstract
features in data automatically by using hierarchical layers In [31] Ai et al. proposed EELSTM technique to train
with complex structure. Currently, it becomes most vibrant the population of LSTM network for forecasting time series
research domain which attracts extraordinary attention from data. In this research they applied the conceptual evolutionary
academia as well as industry. The optimization of DL models cycle for evolving pool of LSTM network in order to achieve
through nature inspired technique is a subject of curiosity more accurate forecast. Farzaneh et al. [32] developed a novel
and debate in computer science, the major application area of method based on GA and Bi-LSTM for melody generation.
this hybrid model in real world applications are autonomous Here GA is used to generate a population of melodies which is
vehicles, biometrics, medical image analysis, data science, further evaluated by Bi-LSTM and those having highest score
machine vision and learning and many more. However, in is selected. Camero et al. [33] explored the mean absolute
this article we are focusing on NIA in deep learning that error (MAE) based random sampling method for multi-layer
is neuro-evolution and parameter optimization etc. DL model architecture and proposed the algorithm based on Evolutionary
is a collective concept of sequence of algorithms and model search for optimizing an RNN using random MAE sampling
which includes but not limited to Long Short-Term Memory as a heuristic search. They showed that proposed technique
Networks (LSTM), Convolutional Neural Networ k(CNN), is training free and low cost for comparing deep RNN
Generative Adversarial Network (GAN), AutoEncoders (AE), architectures and a very useful heuristic method for optimizing
and Echo State Network (ESN) as shown in Fig.2. architecture. In another recent work by Kim and Cho [34] a
model for efficient prediction of residential power demand by a
A. Long-Short-Term-Memory Networks PSO-based CNN-LSTM neural network was proposed. They
LSTM are recurrent deep neural network (DNN) model. used PSO to find optimal CNN-LSTM hyperparameters for
Recently, LSTM network is trained by five nature inspired the household power consumption data set and obtained the
optimizer namely Genetic Algorithm (GA), Gravitational lowest RMSE compared to conventional techniques.

296 10th International Conference on Cloud Computing, Data Science & Engineering (Confluence)
B. Convolutional Neural Networks discriminator population and Frchet Inception Distance (FID)
A CNN is a class of DNN most frequently used for visual at the generator side.They experimented using on MNIST
imaging analysis. Recently, Dixit et al. [35] developed a new dataset and observed that this model solved mode collapse
methodology to find the optimum deep CNN framework for problem partially. This work evidenced that FID score is
texture recognition tasks. In this work They applied the whale a good metric for generator but for discriminator side it
optimization algorithm (WOA) at two different stages. One is needs to evaluate competitive metric. They extended this work
to optimize the values of the filters in the convolution layer and presented an improved model COEGAN in [46] based
and the second one is for optimizing parameters- weight and on neuroevolution and coevolution technique. To implement
biases values of fully connected layers. coevolutionary strategies, COEGAN uses the adversarial
In [36] Strumberger et al. presented a framework called FA- aspect of GAN components in the training algorithm. As
CNN where firefly algorithm (FA) is used to select optimum compare to their earlier work, in this they extended and
CNN architecture on MNIST dataset. For this work they followed the approach of DeepNEAT. The experimental results
optimized number of convolutional and dense layers, number on MNIST and Fashion MNIST datasets shows that proposed
of kernels per layer and the size of the kernel. In another COEGAN achieved a better FID score as well as less number
article an efficient and precise Distributed Denial of Service of layers than DCGAN and random search model. Through
(DDoS) detection method is introduced in [37] using Social this work they tried to demonstrate the efficiency of the
Lion Behavior with CNN. The proposed method of detection evolutionary method because of its mechanism of transferring
is powerful enough to detect DDoS attacks within the lowest weight through generations.
magnitude of traffic attack and acheived 98.2% accuracy on In [47] Wang et al. came with new idea of evolutionary
the data set which is collected in real time using a mininet. GAN (E-GAN) to overcome the problem of instability and
A novel method PSOCNN for searching deep CNN ar- mode collapse during training of GAN and its variants. They
chitecture for image classification proposed in [38]. Beside proposed stable and improved version of GAN and utilized
this, a novel directly encoding technique is also proposed different adversarial training objective as a mutation operator
for comparative analysis of variable length CNNs. Lu et al. for individual generator updation and evolve a population of
[39] presented a new evolutionary neural architecture search generators to adapt the dynamic environment.This model is
NSGA-Net which explores potential CNN architecture based validated on CIFAR10, CelebA and LSUN bedroom dataset
on multiple conflicting objectives namely error metric and and acheived convincing results for image generation.They
computational complexity. The proposed framework achieves suggested video generation and text generation can be con-
3.85% test error rate for CIFAR-10 dataset. Sun et al. [40] sidered for this model. An new evaluation metric based on
also presented an automated evolution of CNN architecture brain inspired network named as neuroscore for GAN was
by applying GA based on ResNet and DenseNet blocks. proposed in [48]. The proposed metric for GAN is consistent
A CNN-based corner detection method for grayscale images for human perception and need lesser number of samples for
is proposed in [41]. In this method,the CNN’s original computation.
processing scheme is altered to include a nonlinear operation
D. Deep Belief Networks
in the image to enhance the contrast of local information.
Further, evolutionary-fuzzy-integral based CNNs (EFI-CNNs) DBN is a class DNN comprised of various layers of both
for age and gender-based facial classification was proposed in directed and undirected edges of the graphic model. Recently,
[42]. The outputs obtained from trained CNNs have been used Sabar et al. [49] proposed an evolutionary hyperheuristic
as input to the EFI and the final fuzzy integral rule has been framework to optimize DBN parameters automatically for the
selected either from the Sugeno or Choquet during testing. problem of image reconstruction. This framework presented
Finally, PSO was used to find optimum fuzzy density values. a high-level approach for automating the operator’s selection
In other recent work, Jones et al. [43] assessed the effec- process and a number of low-level approaches to develop
tiveness and precision of using an evolutionary algorithm to novel parameter settings for DBN while in another research
search the CNNs topology and configuration for automatically Lin et al. [50] developed a model for wind speed forecasting
locating the galaxies in the Avoidance Zone with an accuracy using a DBN with GA called DBNGA. They used GA to
of 91.75%. Beside this, Zhu et al. in [44] examined the determine DBN parameters and both wind speed information
neuro-evolution to automatically design CNN topologies and and multiple weather factors are used to forecast wind
developed a new solution based on Artificial Bee Colony on speed. In [51] Sree et al. introduced a novel framework
the MNIST dataset. called TPSO-DBN that utilizes a new improved PSO namely,
Temperature-Controlled PSO (TPSO) to pre-train a DBN. In
C. Generative Adversarial Networks order to get faster convergence, they attempted to minimize
In DL, the GANs began a revolution. This revolution has the stagnation of particles in the population. The proposed
brought about some significant breakthroughs in research. model is evaluated for Tamil phoneme classification problem
Recently, a new model combining neuroevolution and coevo- with 89.2% classification accuracy.
lution in the GAN training algorithm was presented in [45]. Imbalanced data handling is one of the major challenge for
The proposed model used loss function as a fitness for the classifier. To overcome this issue Ismail and Mohammad in

10th International Conference on Cloud Computing, Data Science & Engineering (Confluence) 297
[52] incorporated the GA and bootstrap sampling into DBN this science will certainly proceed to foster new developments
and tested on imbalance class data. The findings showed a in NIA, with performance and computing effectiveness. The
performance improvement when using Evolutionary DBN with paper presents a short survey on recent applications of NIA
bootstrap sampling. Recently China. Li et al. [53] proposed in DL. In this paper, we expressed our vision for NIC, which
a deep feature-learning strategy to predict short-term traffic calls for a deep reflection on the research directions that the
flows using supervised learning methods with multi-objective world should follow in the future.
PSO (MOPSO). In this MOPSO is used to optimize certain NIA involve setting parameters on their own, and the best
parameters in deep belief networks to obtain traffic flow systematic way of achieving optimal parameter values for
forecast for the next day. NIA remains an open research issue. The addition of NIA
The problem of hyperparameters tuning in Deep Boltzmann in DL therefore constitutes extra parameter values. However,
Machine’s was investigated in [54] using NIC with different the parameter values of the DL models can be decreased as
backgrounds such as fireflies algorithm, variant of harmony some of the parameters can be automatically determined by
search (IHS), Adaptive Inertia Weight PSO (AIWPSO) and the NIA algorithms. This would have a multiplier effect on
variants of differential evolution etc. For validation of pro- the DL in a situation where the NIA parameter values are not
posed techniques they used three publicly available datasets strong enough to deliver great performance. This may have
for binary image reconstruction and also performed the resulted in the model to caught in local minima by reducing
Wilkoxon signed-rank test. Experimental results have shown its performance.
that evolutionary and memory based methods may be more Future research should be on the NIA driven by parame-
appropriate for DBM / DBN hyperparmeter tuning. terless values to eradicate the need for human interference in
E. Autoencoders and Echo State Networks setting parameters values. We hope to be autonomous in future
AE are DNNs that are used to replicate the input at DL models.
the output layer. Recently, convolutional autoencoders with
R EFERENCES
differential evolution (DE) for creating stroke trajectory
writing model and its quality assessment system was presented [1] Y. Nesterov, Lectures on convex optimization. Springer, 2018, vol. 137.
[2] X.-S. Yang, Nature-inspired metaheuristic algorithms. Luniver press,
in [55]. This novel learning framework allows a robot to 2008.
efficiently learn how to write Chinese calligraphy automati- [3] M. Molga and C. Smutnicki, “Test functions for optimization needs,”
cally. Jian-Bo Yu [56] presented a novel stacked denoising Test functions for optimization needs, vol. 101, 2005.
autoencoders (SDAE) algorithm, called manifold regularized [4] A. Darwish, “Bio-inspired computing: Algorithms review, deep analysis,
and the scope of applications,” Future Computing and Informatics
SDAE (MRSDAE) with PSO for gearbox fault detection in Journal, vol. 3, no. 2, pp. 231–246, 2018.
which manifold regularization and feature engineering are [5] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” nature, vol. 521,
smoothly integrated into the deep network. no. 7553, p. 436, 2015.
[6] A. Lopez-Rincon, A. Tonda, M. Elati, O. Schwander, B. Piwowarski,
In another recent work, Silhan et al. [57] proposed a novel and P. Gallinari, “Evolutionary optimization of convolutional neural
SAE and presented automatic and adaptive hyperparameters networks for cancer mirna biomarkers classification,” Applied Soft
optimization where layer wise training is done. This approach Computing, vol. 65, pp. 91–100, 2018.
[7] X.-W. Chen and X. Lin, “Big data deep learning: challenges and
is validated on MNIST, Fashion MNIST, and the Year perspectives,” IEEE access, vol. 2, pp. 514–525, 2014.
Prediction Million Song Database and showed promising [8] R. Garcı́a-Ródenas, L. J. Linares, and J. A. López-Gómez, “A
results. Charte et al. in [58] showed how to automate AE memetic chaotic gravitational search algorithm for unconstrained global
optimization problems,” Applied Soft Computing, vol. 79, pp. 14–29,
architecture using evolutionary methods. In this approach they 2019.
initially proposed a coding to incorporate the AE configuration [9] H. Bouchekara, “Solution of the optimal power flow problem consider-
in a chromosome. Then, two evolutionary possibilities are ing security constraints using an improved chaotic electromagnetic field
optimization algorithm,” Neural Computing and Applications, pp. 1–21,
contrasted against search space. In another research by 2019.
Rodriguez et al. [59] used genetic programming (GP) to design [10] G. I. Sayed, A. Tharwat, and A. E. Hassanien, “Chaotic dragonfly
the AE structures for high dimensional data representation algorithm: an improved metaheuristic algorithm for feature selection,”
Applied Intelligence, vol. 49, no. 1, pp. 188–205, 2019.
learning .It also includes a partitioning system with high- [11] H. A. Abdulwahab, A. Noraziah, A. A. Alsewari, and S. Q. Salih, “An
dimensional input representations for distributed and online enhanced version of black hole algorithm via levy flight for optimization
learning. and data clustering problems,” IEEE Access, 2019.
[12] A. Baykasoğlu and M. E. Şenol, “Weighted superposition attraction algo-
Further, Chouikhi et al. in [60] introduced multi layered rithm for combinatorial optimization,” Expert Systems with Applications,
echo state neteork (ML-ESN) for the first time as an vol. 138, p. 112792, 2019.
autoencoder to extract the new features from the given dataset. [13] R. Logesh, V. Subramaniyaswamy, V. Vijayakumar, X.-Z. Gao, and G.-
G. Wang, “Hybrid bio-inspired user clustering for the generation of
A novel and effective bi level evolutionary approach using diversified recommendations,” Neural Computing and Applications, pp.
MOPSO and PSO is experimented for both noisy and noise 1–20, 2019.
free environment. [14] C.-B. Cheng, H.-S. Shih, and E. S. Lee, “Metaheuristics for multi-
level optimization,” in Fuzzy and Multi-Level Decision Making: Soft
IV. CONCLUSIONS Computing Approaches. Springer, 2019, pp. 171–188.
[15] M. A. Tawhid and V. Savsani, “Multi-objective sine-cosine algorithm
Nature is incredibly interesting and yet to be interpreted full (mo-sca) for multi-objective engineering design problems,” Neural
of intriguing occurrences. Undiscovered paradigms underneath Computing and Applications, vol. 31, no. 2, pp. 915–929, 2019.

298 10th International Conference on Cloud Computing, Data Science & Engineering (Confluence)
[16] W. Long, J. Jiao, X. Liang, S. Cai, and M. Xu, “A random opposition- [39] Z. Lu, I. Whalen, V. Boddeti, Y. Dhebar, K. Deb, E. Goodman, and
based learning grey wolf optimizer,” IEEE Access, vol. 7, pp. 113 810– W. Banzhaf, “Nsga-net: neural architecture search using multi-objective
113 825, 2019. genetic algorithm,” in Proceedings of the Genetic and Evolutionary
[17] A. A. Heidari, S. Mirjalili, H. Faris, I. Aljarah, M. Mafarja, and Computation Conference. ACM, 2019, pp. 419–427.
H. Chen, “Harris hawks optimization: Algorithm and applications,” [40] Y. Sun, B. Xue, M. Zhang, and G. G. Yen, “Completely automated
Future Generation Computer Systems, vol. 97, pp. 849–872, 2019. cnn architecture design based on blocks,” IEEE transactions on neural
[18] D. Tian, X. Zhao, and Z. Shi, “Chaotic particle swarm optimization networks and learning systems, 2019.
with sigmoid-based acceleration coefficients for numerical function [41] E. Cuevas, M.-A. Dı́az-Cortes, and E. Mezura-Montes, “Corner detection
optimization,” Swarm and Evolutionary Computation, p. 100573, 2019. of intensity images with cellular neural networks (cnn) and evolutionary
[19] X.-S. Yang and X.-S. He, “Mathematical analysis of algorithms: Part i,” techniques,” Neurocomputing, vol. 347, pp. 82–93, 2019.
in Mathematical Foundations of Nature-Inspired Algorithms. Springer, [42] C.-J. Lin, C.-H. Lin, C.-C. Sun, and S.-H. Wang, “Evolutionary-
2019, pp. 59–73. fuzzy-integral-based convolutional neural networks for facial image
[20] M. M. Mafarja and S. Mirjalili, “Hybrid binary ant lion optimizer with classification,” Electronics, vol. 8, no. 9, p. 997, 2019.
rough set and approximate entropy reducts for feature selection,” Soft [43] D. Jones and G. Nitschke, “Evolutionary deep learning to identify
Computing, vol. 23, no. 15, pp. 6249–6265, 2019. galaxies in the zone of avoidance,” arXiv preprint arXiv:1903.07461,
[21] C. Huang, Y. Li, and X. Yao, “A survey of automatic parameter 2019.
tuning methods for metaheuristics,” IEEE Transactions on Evolutionary [44] W. Zhu, W. Yeh, J. Chen, D. Chen, A. Li, and Y. Lin, “Evolutionary
Computation, 2019. convolutional neural networks using abc,” in Proceedings of the 2019
[22] M. Isiet and M. Gadala, “Self-adapting control parameters in particle 11th International Conference on Machine Learning and Computing.
swarm optimization,” Applied Soft Computing, vol. 83, p. 105653, 2019. ACM, 2019, pp. 156–162.
[23] R. Vafashoar and M. R. Meybodi, “Cellular learning automata based [45] V. Costa, N. Lourenço, and P. Machado, “Coevolution of generative
bare bones pso with maximum likelihood rotated mutations,” Swarm adversarial networks,” in International Conference on the Applications
and evolutionary computation, vol. 44, pp. 680–694, 2019. of Evolutionary Computation (Part of EvoStar). Springer, 2019, pp.
473–487.
[24] Y. Wang, H. Liu, Z. Yu, and L. Tu, “An improved artificial neural
[46] V. Costa, N. Lourenço, J. Correia, and P. Machado, “Coegan: evaluating
network based on human-behaviour particle swarm optimization and
the coevolution effect in generative adversarial networks,” in Proceed-
cellular automata,” Expert Systems with Applications, vol. 140, p.
ings of the Genetic and Evolutionary Computation Conference. ACM,
112862, 2020.
2019, pp. 374–382.
[25] P. Jiang, Q. Zhou, J. Liu, and Y. Cheng, “A three-stage surrogate model
[47] C. Wang, C. Xu, X. Yao, and D. Tao, “Evolutionary generative
assisted multi-objective genetic algorithm for computationally expensive
adversarial networks,” IEEE Transactions on Evolutionary Computation,
problems,” in 2019 IEEE Congress on Evolutionary Computation (CEC).
2019.
IEEE, 2019, pp. 1680–1687.
[48] Z. Wang, Q. She, A. F. Smeaton, T. E. Ward, and G. Healy, “Neuroscore:
[26] Y. Du, T. Wang, B. Xin, L. Wang, Y. Chen, and L. Xing, “A data- A brain-inspired evaluation metric for generative adversarial networks,”
driven parallel scheduling approach for multiple agile earth observation arXiv preprint arXiv:1905.04243, 2019.
satellites,” IEEE Transactions on Evolutionary Computation, 2019. [49] N. R. Sabar, A. Turky, A. Song, and A. Sattar, “An evolutionary hyper-
[27] A. Santiago, B. Dorronsoro, A. J. Nebro, J. J. Durillo, O. Castillo, and heuristic to optimise deep belief networks for image reconstruction,”
H. J. Fraire, “A novel multi-objective evolutionary algorithm with fuzzy Applied Soft Computing, p. 105510, 2019.
logic based adaptive selection of operators: Fame,” Information Sciences, [50] K.-P. Lin, P.-F. Pai, and Y.-J. Ting, “Deep belief networks with genetic
vol. 471, pp. 233–251, 2019. algorithms in forecasting wind speed,” IEEE Access, vol. 7, pp. 99 244–
[28] C. Yue, B. Qu, K. Yu, J. Liang, and X. Li, “A novel scalable test 99 253, 2019.
problem suite for multimodal multiobjective optimization,” Swarm and [51] B. L. Sree and M. Vijaya, “Temperature controlled pso on optimizing
Evolutionary Computation, vol. 48, pp. 62–71, 2019. the dbn parameters for phoneme classification,” International Journal of
[29] A. Q. B. Silva, “Predicting cervical cancer with metaheuristic optimizers Speech Technology, vol. 22, no. 1, pp. 143–156, 2019.
for training lstm,” in International Conference on Computational [52] A. R. Ismail, O. A. Mohammad et al., “Evolutionary deep belief
Science. Springer, 2019, pp. 642–655. networks with bootstrap sampling for imbalanced class datasets,”
[30] Y. Li, Z. Zhu, D. Kong, H. Han, and Y. Zhao, “Ea-lstm: Evolutionary International Journal of Advances in Intelligent Informatics, vol. 5,
attention-based lstm for time series prediction,” Knowledge-Based no. 2, pp. 123–136, 2019.
Systems, 2019. [53] L. Li, L. Qin, X. Qu, J. Zhang, Y. Wang, and B. Ran, “Day-ahead
[31] S. Ai, A. Chakravorty, and C. Rong, “Evolutionary ensemble lstm based traffic flow forecasting based on a deep belief network optimized by the
household peak demand prediction,” in 2019 International Conference multi-objective particle swarm algorithm,” Knowledge-Based Systems,
on Artificial Intelligence in Information and Communication (ICAIIC). vol. 172, pp. 1–14, 2019.
IEEE, 2019, pp. 1–6. [54] L. A. Passos and J. P. Papa, “A metaheuristic-driven approach to fine-
[32] M. Farzaneh and R. M. Toroghi, “Melody generation using an interactive tune deep boltzmann machines,” Applied Soft Computing, p. 105717,
evolutionary algorithm,” arXiv preprint arXiv:1907.04258, 2019. 2019.
[33] A. Camero, J. Toutouh, and E. Alba, “A specialized evolutionary strategy [55] X. Gao, C. Zhou, F. Chao, L. Yang, C.-M. Lin, T. Xu, C. Shang,
using mean absolute error random sampling to design recurrent neural and Q. Shen, “A data-driven robotic chinese calligraphy system using
networks,” arXiv preprint arXiv:1909.02425, 2019. convolutional auto-encoder and differential evolution,” Knowledge-
[34] T.-Y. Kim and S.-B. Cho, “Particle swarm optimization-based cnn-lstm Based Systems, 2019.
networks for forecasting energy consumption,” in 2019 IEEE Congress [56] J.-B. Yu, “Evolutionary manifold regularized stacked denoising autoen-
on Evolutionary Computation (CEC). IEEE, 2019, pp. 1510–1516. coders for gearbox fault diagnosis,” Knowledge-Based Systems, vol. 178,
[35] U. Dixit, A. Mishra, A. Shukla, and R. Tiwari, “Texture classification pp. 111–122, 2019.
using convolutional neural network optimized with whale optimization [57] T. Silhan, S. Oehmcke, and O. Kramer, “Evolution of stacked autoen-
algorithm,” SN Applied Sciences, vol. 1, no. 6, p. 655, 2019. coders,” in 2019 IEEE Congress on Evolutionary Computation (CEC).
[36] I. Strumberger, E. Tuba, N. Bacanin, M. Zivkovic, M. Beko, and IEEE, 2019, pp. 823–830.
M. Tuba, “Designing convolutional neural network architecture by the [58] F. Charte, A. J. Rivera, F. Martı́nez, and M. J. del Jesus, “Automating
firefly algorithm,” in 2019 International Young Engineers Forum (YEF- autoencoder architecture configuration: An evolutionary approach,” in
ECE). IEEE, 2019, pp. 59–65. International Work-Conference on the Interplay Between Natural and
[37] D. Arivudainambi, V. K. KA, and S. S. Chakkaravarthy, “Lion ids: A Artificial Computation. Springer, 2019, pp. 339–349.
meta-heuristics approach to detect ddos attacks against software-defined [59] L. Rodriguez-Coayahuitl, A. Morales-Reyes, and H. J. Escalante,
networks,” Neural Computing and Applications, vol. 31, no. 5, pp. 1491– “Evolving autoencoding structures through genetic programming,”
1501, 2019. Genetic Programming and Evolvable Machines, pp. 1–28, 2019.
[38] F. E. F. Junior and G. G. Yen, “Particle swarm optimization of deep [60] N. Chouikhi, B. Ammar, A. Hussain, and A. M. Alimi, “Bi-level multi-
neural networks architectures for image classification,” Swarm and objective evolution of a multi-layered echo-state network autoencoder
Evolutionary Computation, 2019. for data representations,” Neurocomputing, vol. 341, pp. 195–211, 2019.

10th International Conference on Cloud Computing, Data Science & Engineering (Confluence) 299

You might also like