A Systematic Assessment of Genetic Algorithm (GA) in Optimizing Machine

Learning Model: A Case Study from Building Science

2022 IEEE International Conference on Industrial Engineering and Engineering Management (IEEM) | 978-1-6654-8687-3/22/$31.00 ©2022 IEEE | DOI: 10.1109/IEEM55944.2022.9989719

A. Ali1, R. Jayaraman1, E. Azar2, A. Sleptchenko1

1
Department of Engineering Systems and Management, Khalifa University, Abu Dhabi, United Arab Emirates
2
Department of Civil and Environmental Engineering, Carleton University, Ottawa, ON, Canada
([email protected])

Abstract - Machine learning (ML) algorithms are such as first-order optimizers (Batch gradient
techniques that allow computers to learn from the data descent, AdaMax, AdaDelta, among others), second-order
without being explicitly programmed. ML techniques consist optimizers (quasi-Newton’s method, conjugate gradient
of hyperparameters that typically influence prediction descent, among others), swarm intelligence optimizers
accuracy, hence requiring tuning. In this study, we (swarm optimizers, grey wolf optimizers, among others).
systematically evaluate the performance of the genetic
algorithm (GA) technique in tuning ML hyperparameters
Other methods such as greedy search, heuristic algorithms,
compared to three other common tuning techniques i.e. grid reinforcement learning, and evolutionary algorithms are
search (GS), random search (RS), and bayesian optimization also being applied [6]. As mentioned, GS or sometimes
(BO). While previous studies explored the potential of referred to as exhaustive search has been used in ML
metaheuristics techniques such as GA in tuning ML models, hyperparameter optimization since the 1990s [4]. It is
a systematic comparison with other commonly mentioned simply selecting and testing all possible combinations from
techniques is currently lacking. Results indicate that GA the defined range of hyperparameters, hence a brute force
slightly outperformed other methods in terms of optimality technique. GS has been applied in several ML
due to its ability to pick any continuous value within the hyperparameter tuning studies such as [3], [2]. Although
range. However, apart from GS which took the longest, it was
observed that GA is quite a time inefficient compared to RS
GS can guarantee optimality within the defined range of
and BO which were able to find a solution close to the GA values (defined grid), it can also be observed that this
within a shorter time (GA – 149 minutes, RS – 88 minutes, BO technique is not time efficient. For example, 6
– 105 minutes, GS – 756 minutes). hyperparameters with 10 values will require 106 iterations
which are computationally expensive and time-costly [7],
Keywords – Genetic algorithm, Hyperparameter, [8].
Machine learning,
Due to GS time limitations, another technique known as a
I. INTRODUCTION random search (RS) is introduced. RS selects and tests
random combinations from a set of possible combinations
A. Background until a certain stopping criterion is met [4]. The stopping
criteria include reaching a threshold prediction accuracy
Generally, machine learning (ML) models can be regarded and reaching a maximum number of iterations i.e. number
as optimization problems as it is looking for the best fit of of randomly tested combinations, among others. Although
independent variables to the target variable, some such in some problems, RS can perform as well as a GS [9], it is
examples are mentioned in [1]. However, a more common not always the case. This typically results in a trade-off
application of optimization techniques in machine learning between the two techniques. Unlike GS and RS, another
is in tuning the ML model hyperparameters. Machine technique known as Bayesian optimization (BO) is a
learning model performance is typically influenced by model-based technique that approximates the true
some of the parameters known as hyperparameters. For objective function using a surrogate model such as the
example, any tree-based model such as random forest (RF), Gaussian process. Rather than exploring randomly in the
gradient boosting (GB), and extreme gradient boosting search space, this technique has a mechanism to explore the
(XGB) are influenced by hyperparameters such as the search space areas that will improve the results [10]. Just
number of trees generated, maximum size of the tree, like RS, BO is also controlled by the number of iterations,
contribution of each tree in the overall model, among which reflects the number of hyperparameter combinations
others. [2]. Likewise, artificial neural network (ANN) tested. Additionally, metaheuristic techniques such as
models are influenced by the number of hidden layers. genetic algorithm (GA) is another optimization technique
Hence, finding the best combination of hyperparameters is that is widely applied in tuning ML models such as deep
quite crucial in providing a better predictive performance learning, tree-based models, artificial neural networks
of models. Several optimization techniques are used in (ANN), Support Vector Machine (SVM), Naïve Bayes, K-
tuning machine learning (ML) hyperparameters. nearest neighbors (KNN), polynomial, and logistic
Traditional methods such as grid search (GS) and random regression [11], [12], [13], [14], [15], [16]. Conceptually,
search (RS) are widely used [2], [3], [4]. Additionally, GA starts with the initial population and keeps on selecting
other techniques are also being applied, for example, [5]
mentions techniques that are tested in tuning deep learning

0384
Authorized licensed use limited to: Arizona State University. Downloaded on April 09,2025 at 03:23:06 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of The 2022 IEEE International Conference on Industrial Engineering and Engineering Management (IEEM)

the fitter individual for next-generation evaluations and TABLE I

hoping to get the best individual at convergence. DATASET VARIABLES
INDEPENDENT VARIABLES DEPENDENT VARIABLES
Relative compactness Heating load
While ML is applicable in several domains, one area that is Surface area Cooling load
observed to apply ML is in the building science domain. Wall area
[17],[28],[29]. Currently, there is plenty of research in Roof area
building science covering stages of the building life cycle Overall height
Orientation
that are involving machine learning techniques using ML Glazing area
methods such as support vector machines, artificial neural Glazing area distribution
networks, etc. [17]. Knowing which optimization
technique is suitable for modeling building science datasets TABLE II
could be important and catalyze the knowledge discovery GRADIENT BOOSTING HYPERPARAMETERS [2]
process in building science. HYPERPARA DESCRIPTION RANGE
METER
max depth Size of the tree 3, 4, 5, 6, 8, 9, 10, 12, 15
B. Motivation n estimators Number of trees 100, 150, 200, 250, 300,
400, 500, 600, 800, 1000,
Taking into consideration the time and optimality trade-off, 1200
plenty of research is going on exploring novel, efficient and min samples Samples required in 2, 3, 4, 5, 6, 8,10, 15
better optimization algorithms [18] but it is still worth it split the leaf to split
learning rate Contribution of tree 0.01, 0.02, 0.05, 0.1, 0.2,
and interesting to compare the performance of the existing in the final 0.25, 0.3, 0.4, 0.5
optimization techniques when applied to certain problems. prediction
For example, [10] assessed the performance of GS, RS, and
BO in a small and medium-sized medical dataset. While TABLE III
[19] compared GS, RS, spectral analysis, BO, and CMA- GA PARAMETERS
ES techniques and observed that spectral analysis is more PARAMETER RANGE
capable in parallel ML parameter optimization, while BO Population 25, 50
is efficient in sequential ML parameter optimization and Crossover probability 0.1, 0.5, 0.9
Mutation probability 0.01, 0.5, 0.9
CMA-ES performed well in larger datasets. Reference [16] Generation 50
did not focus on the comparison of the algorithm but
instead acknowledged the existence of methods such as BO The dataset was generated using Ecotect in which 12
and evolutionary algorithms that are overcoming the curse buildings were simulated using several variations such as
of dimensionality issue, which is quite a challenge in the surface areas, dimensions, glazing areas, etc with their
machine learning domain. To the best of the author’s respective heating and cooling loads. A total of 768
knowledge, there is a lack of systematic comparison in instances comprising 8 independent variables and 2
terms of optimality and time cost between traditional dependent variables as seen in Table. I. The energy-
techniques such as GS, RS, a model-based technique such efficient dataset is covering one of the main challenges in
as BO, and meta-heuristic techniques such as GA in building science which is energy consumption in buildings.
optimizing ML hyperparameters. According to [21], 40% of the total energy is used in
residential and commercial buildings. Limiting the scope
C. Objective
of this study, only heating load is considered which is of
Hence, the main contribution of this paper is to assess the continuous type and hence regression machine learning
performance of the genetic algorithm in tuning machine problem.
learning model hyperparameters by benchmarking with
other traditional optimization techniques such as grid B. Machine learning (ML)
search, random search, and bayesian optimization.
Machine learning is simply the process of allowing the
II. METHODOLOGY computer to train itself and learn to make a decision based
on the dataset without program instruction from human
This section presents the research methodology using the intervention. Several techniques have been proposed over
following four (4) sections; 2.1 presents the dataset used, the years, one of them is Gradient Boosting (GB). GB is
2.2 presents the machine learning algorithm used in the one of the tree-based ensemble algorithms that iteratively
study, 2.3 presents the GA method, and finally, 2.4 presents generate trees and later combine the trees to form a single
the experiment design which explains the research strong predictor. The final prediction is made by the
approach. weighted contribution from each tree. In the GB model,
several hyperparameters can be tuned to optimize the
A. Dataset performance of the model. Some of those hyperparameters
The dataset used in this study is referred to as an energy- are affecting individual trees and some are affecting the
efficient dataset in the UCI repository website created whole model as seen in Table. II.
by [20].

0385
Authorized licensed use limited to: Arizona State University. Downloaded on April 09,2025 at 03:23:06 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of The 2022 IEEE International Conference on Industrial Engineering and Engineering Management (IEEM)

C. Genetic Algorithm (GA)

GA is one of the branches within evolutionary

algorithms that inherits the concept of biological genes. A
gene is a part of a chromosome (or in the GA world known
as an individual), representing a feature and the value
within each position indicates the value of that particular
feature [22]. Hence, an individual is made of a collection
of features and their respective values just like the
collection of genes making up a chromosome. With the
notion of “survival of the fittest” in evolutionary
algorithms, GA is accessing the fitness of each individual
by obtaining the objective values of each individual in a
population and only those that performed well according to
the problem at hand are having higher probabilities of
Fig. 1. Methodology framework
participating in the next generations of individuals. The
selection of individuals, tournament selection, and roulette
wheel techniques are typically used. Most techniques have optimization, the optimal combinations are used as default
elements of randomness to allow a lesser fit individual to parameters of the GA algorithm in phase 3. Phase 3 takes
be selected as well as to ensure diversity and avoid early both the defined model and hyperparameter range as inputs
convergence [23]. After the selection process is over, a list and apply in each of the mentioned optimization
of selected individuals are either combined (i.e. crossover) techniques. With the defined hyperparameter range, GS
or slightly changed (i.e. mutation) to generate the next needs 7920 iterations while both RS and BO were tested
generation of individual that will be tested for fitness again using a range of iterations [50, 100, 200, 500, 1000] to get
and expecting a better performance. A crossover rate also more insight into the performance. The objective function
known as crossover probability defines the number of refers to minimizing the mean squared error as explained
occurrences of crossover in a generation while mutation in Section. 2 of this paper. One thing to note is that the
rate or mutation probability defines the number of minimized mean squared error is a 10-fold cross-validation
chromosomes to be mutated in a generation, both with a error and not a testing error. While each optimization
similar range of 0 to 1 [24]. The performance of any GA technique requires a different number of iterations
depends on many GA parameters but the four most depending on its nature, a comparison between
common and important parameters are crossover optimization techniques will be based on time and
probability, mutation probability, number of generations, closeness to the optimal solution to ensure a fair
and the number of individuals (population) in each of the comparison of analysis in phase 4. GS will only be used to
generations [24]. Hence, parameter tuning is needed to provide a benchmark of the optimal solution by searching
obtain optimal parameters. Table III shows the selected over all possible combinations. The remaining
parameters with their respective ranges. It can be noted that optimization methods will be benchmarked against the GS
the ranges applied are quite narrow, this is due to both results to observe how close the solution is to the optimal
computational resource limitations. and how fast it reached the optimality. All models were
developed in a python environment using the scikit-
D. Experiment design learn library [25], hyperOpt library (for BO) [26], and
DEAP library (for GA optimization) [27]. The experiment
The experiment design is composed of four (4) main was conducted using Intel core i7 processor at 1.9GHz with
phases. The initial phase is the data pre-process phase 16GB RAM.
which; first split the dataset into training and testing sets
(80% for training and 20% for testing). Phase 2 is the
model initializing phase, in which the GB model was III. RESULTS
initialized with 4 main hyperparameters mentioned in
Table. II to be optimized. During phase 2, other model- This section presents and analyzes; first, the results
based parameters are fixed as well such as cross-validation obtained during the GA tuning phase in Section. 3.1 and
fold (10 folds) and evaluation metric (mean-squared error the results of the machine learning hyperparameter tuning
- MSE). The initialized model is going to be optimized phase in Section. 3.2.
using several optimization techniques mentioned in phase
3. In phase 2, GA parameter optimization is conducted as A. GA parameter tuning
well. For GA, based on the defined GA parameter ranges
In this section, the discussion focuses on the results
seen in Table. III, there is 18 parameter combination that
obtained in tuning the GA parameters. As explained in
needs to run over 50 generations each. After GA parameter
Section. 3, GA parameters tuning needs to be performed to
obtain the optimal set of parameters of the GA algorithm.

0386
Authorized licensed use limited to: Arizona State University. Downloaded on April 09,2025 at 03:23:06 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of The 2022 IEEE International Conference on Industrial Engineering and Engineering Management (IEEM)

B. ML hyperparameter tuning

To start with, GS which is known to reach optimality

within the defined grid obtained an MSE of 0.1146. The
obtained optimal MSE is slightly higher compared to the
MSE obtained by [2] which used a similar dataset and
similar hyperparameter grid values. This is mainly due to
the variation of the fixed seed for the dataset used in the
model. This paper used a fixed seed of 5 while in [2] the
seed number was not mentioned.

Fig. 2. GA parameter tuning results

TABLE IV
GA PARAMETERS COMBINATIONS (* refers to
optimal set)
SET NO. POP. SIZE CP MP GEN.
1 25 0.1 0.01 50
2 25 0.1 0.5 50
3 25 0.1 0.9 50
4 25 0.5 0.01 50
5 25 0.5 0.5 50
6 25 0.5 0.9 50
7 25 0.9 0.01 50
8 25 0.9 0.5 50 Fig. 3. RS results compared to GS
9 25 0.9 0.9 50
10 50 0.1 0.01 50 For the RS, several iterations were evaluated i.e. 50, 100,
11 50 0.1 0.5 50
12 50 0.1 0.9 50 200, 500, and 1000 iterations. A significant observation
13 50 0.5 0.01 50 from Fig. 3, shows that none of the RS results were able to
14* 50 0.5 0.5 50 be better than the GS result. As it is obvious that the
15 50 0.5 0.9 50 optimal MSE obtained from RS can only be as good as GS
16 50 0.9 0.01 50
17 50 0.9 0.5 50 but never better. This is because RS is picking a set of
18 50 0.9 0.9 50 parameters randomly from the grid instead of iterating
through all of them as in GS, so there is a chance that the
The list of parameters and their respective ranges are listed randomly picked combination to be the optimal one within
in Table. III. With 50 generations in each set of parameters, the specified range. It was also observed that, with a higher
a total of 18 combinations or sets of combinations were number of iterations, even better optimal MSE was
tested. The optimal parameters are then fixed when tuning obtained. However, this may not always be the case due to
the ML hyperparameters. All tested combinations (set of the randomness within the RS.
parameters) are listed in Table. IV. From Fig. 2, it can be
noted that set 14 obtained minimum MSE (0.11417), hence
the optimal set of parameters. The results highlight the
necessity of tuning GA parameters as clearly observed
different values of parameters yielded different MSE
values. Some sets obtained MSE values close to the
optimal value, for example, set 3, 12, and 16 while other
sets obtained higher MSEs such as sets 1, 8, and 13.

Set 14 had a population size of 50 with 50 generations, both

crossover probability and mutation probability of 0.5. An
important thing to note here is that even with fixed GA
parameters; the stochastic nature within GA may result in Fig. 4. BO results compared to GS
different results, hence replicability may be almost
impossible. Some of the stochastic characteristics within With BO, a similar number of iterations used in the RS
GA include the initial selection of populations, the were tested and the results obtained were again compared
randomness of crossover, and the randomness of mutation to obtained GS results as seen in Fig. 4. It can be observed
within individuals. from Fig. 4 that, some of the trials were able to get close to
the GS solution. Considering time limitations, BO can get
solutions as close as GS solution within fewer iterations
(1000 iterations in this paper) compared to 7920 iterations

0387
Authorized licensed use limited to: Arizona State University. Downloaded on April 09,2025 at 03:23:06 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of The 2022 IEEE International Conference on Industrial Engineering and Engineering Management (IEEM)

from GS. Unlike RS, one of the advantages of BO is its

capability of finding even better solutions compared to GS
because it picks randomly any real number specified
between the lower bound and upper bound of the non-
integer hyperparameter while GS picks only specified
values within the range.

Moving to the GA technique, there are quite interesting

observations as well related to the optimality of the solution
as well as run time. As mentioned in Section. 3.1, the
randomness observed in generating the initial population is
playing an important role in getting towards better
solutions. For example, during the GA optimization phase,
the optimal solution was 0.1142 better than the GS result Fig. 5. Genetic algorithm compared to GS
as seen in Fig. 5. But when tuning the ML hyperparameter
with the obtained GA parameters, an even better MSE of
0.1128 was recorded but it could have been worse as well.
This is caused by not only the randomness in generating the
initial population but also the populations in the
consecutive generations based on the stochastic
mechanism in performing crossover and mutations. Hence,
the randomness within the GA method might lead to some
results variations. Just like BO, GA can outperform GS due
to its ability to select any real value between the lower and
upper bound for each of the non-integer hyperparameters.
Even though GA is capable of obtaining better solutions
compared to GS but it is observed to take longer running
time compared to RS, and BO as seen in Fig. 6. The optimal Fig. 6. Optimal solution vs time taken
MSE and optimal hyperparameters are seen in Table. V.
TABLE V
Even though GA was able to obtain numerically the best
OPTIMIZATION METHODS PERFORMANCE
solution but it can be observed that the difference with
SUMMARY (MD refers to max depth, MSS refers to min
other solutions is quite minor (0.1128 – 0.1164 - the
samples split, NE refers to n estimators, and LR refers to
difference is observed from the third decimal place).
learning rate)
IV. DISCUSSION AND CONCLUSION METHOD MSE TIME MD MSS NE LR
GA 0.1128 149.46 5 5 794 0.08
The main aim of this paper was to conduct a systematic
assessment of GA in optimizing machine learning models BO 0.116 105.32 5 2 800 0.07
compared to traditional techniques such as GS and RS, and RS 0.1164 51.64 5 4 300 0.25
also model-based techniques such as BO using a dataset GS 0.1146 756.97 10 15 100 0.3
from the building science domain. With the study approach
proposed in Section. 2, it was observed that GA can get a
better optimal solution, and all other techniques obtained It is quite important to point out that these results might be
an optimal solution a little less compared to GA. Another sensitive to the type of dataset used or even the machine
main observation is that, though GA got the best optimal learning algorithm used. Hence for future studies, the
solution, it took longer compared to RS and BO while GS author is recommending the following, firstly, conducting
took the longest. It was also observed that the stochastic a similar study on a bigger dataset and observing the results
elements within the GA mechanics such as the random compared to the small dataset used in this study. Secondly,
generation of the initial population, or crossover due to stochastic nature of RS, BO, and even GA, the
probability are also causing variations in the solutions even results obtained may be influenced by randomness, hence
with fixed GA parameters. On a positive note, GA just like future studies should consider running several runs to
BO can find even better results compared to GS since it can improve the interpretability of results. Finally, other types
pick any continuous value between the upper and lower of machine learning algorithms and evaluation metrics
bound of a non-integer hyperparameter, hence there is a should be evaluated as well.
chance of picking better combinations compared to the grid
and RS which are strictly controlled by the defined range.

0388
Authorized licensed use limited to: Arizona State University. Downloaded on April 09,2025 at 03:23:06 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of The 2022 IEEE International Conference on Industrial Engineering and Engineering Management (IEEM)

REFERENCES [16] H. H. Elmousalami and M. Elaskary, “Drilling stuck pipe

classification and mitigation in the Gulf of Suez oil fields
[1] C. Gambella, B. Ghaddar, and J. Naoum-Sawaya, using artificial intelligence,” Journal of Petroleum
“Optimization problems for machine learning: A survey,” Exploration and Production Technology, pp. 1–14, 2020.
European Journal of Operational Research, aug 2020. [17] T. Hong, Z. Wang, X. Luo, and W. Zhang, “State-of-the-art
[Online]. Available: https: on research and applications of machine learning in the
//doi.org/10.1016%2Fj.ejor.2020.08.045 building life cycle,” Energy and Buildings, vol. 212, p.
[2] S. Papadopoulos, E. Azar, W.-L. Woon, and C. E. 109831, 2020.
Kontokosta, “Evaluation of tree-based ensemble learning [18] Y. Huan, F. Wu, M. Basios, L. Kanthan, L. Li, and B. Xu,
algorithms for building energy performance estimation,” “IEO: Intelligent Evolutionary Optimisation for
Journal of Building Performance Simulation, vol. 11, no. 3, Hyperparameter Tuning,” arXiv preprint
pp. 322–332, 2017. arXiv:2009.06390, 2020.
[3] M. Geitle and R. Olsson, “A New Baseline for Automated [19] W. Seubring and D. Timmerman, “Selecting the optimal
Hyper- Parameter Optimization,” in Machine Learning hyperparameter optimization method: a comparison of
Optimization, and Data Science. Springer International methods,” 16th SC@ RUG 2018-2019, p. 59, 2019.
Publishing, 2019, pp. 521–530. [Online]. Available: [20] A. Tsanas and A. Xifara, “Accurate quantitative estimation
https://doi.org/10.1007%2F978-3-030-37599-7 43 of energy performance of residential buildings using
[4] G. H. John, “Cross-validated C4. 5: Using error estimation statistical machine learning tools,” Energy and Buildings,
for automatic parameter selection,” Training, vol. 3, p. 1, vol. 49, pp. 560–567, 2012.
1994. [21] S. S. Gilan and B. Dilkina, “Sustainable building design: A
[5] C. E. Nwankpa, “Advances in optimisation algorithms and challenge at the intersection of machine learning and design
techniques for deep learning,” Advances in Science, optimization,” in AAAI Workshop: Computational
Technology and Engineering Systems Journal, vol. 5, no. 5, Sustainability, 2015.
pp. 563–577, 2020. [22] S. Sivanandam and S. Deepa, “Genetic algorithms,” in
[6] M. Javaheripi, M. Samragh, and F. Koushanfar, “Peeking Introduction to genetic algorithms. Springer, 2008, pp. 15–
Into the Black Box: A Tutorial on Automated Design 37.
Optimization and Parameter Search,” IEEE Solid-State [23] M. Kumar, M. Husain, N. Upreti, and D. Gupta, “Genetic
Circuits Magazine, vol. 11, no. 4, pp. 23–28, 2019 algorithm: Review and application,” Available at SSRN
[7] A. H. Victoria and G. Maragatham, “Automatic tuning of 3529843, 2010.
hyperparameters using BO,” 2020. [24] A. Hassanat, K. Almohammadi, E. Alkafaween, E.
[8] M. Claesen, J. Simm, D. Popovic, Y. Moreau, and B. De Abunawas, A. Hammouri, and V. Prasath, “Choosing
Moor, “Easy hyperparameter search using optunity,” arXiv mutation and crossover ratios for genetic algorithms—a
preprint arXiv:1412.1114, 2014. review with a new dynamic approach,” Information,
[9] P. Probst, M. N. Wright, and A.-L. Boulesteix, vol. 10, no. 12, p. 390, 2019.
“Hyperparameters and tuning strategies for random forest,” [25] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B.
Wiley Interdisciplinary Reviews: Data Mining and Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss,
Knowledge Discovery, vol. 9, no. 3, p. e1301, 2019. V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M.
[10] V. J. Kadam and S. M. Jadhav, “Performance analysis of Brucher, M. Perrot, and E. Duchesnay, “Scikit-learn:
hyperparameter optimization methods for ensemble Machine Learning in Python,” Journal of Machine Learning
learning with small and medium sized medical datasets,” Research, vol. 12, pp. 2825–2830, 2011.
Journal of Discrete Mathematical Sciences and [26] J. Bergstra, D. Yamins, and D. Cox, “Making a science of
Cryptography, vol. 23, no. 1, pp. 115–123, 2020. model search: Hyperparameter optimization in hundreds of
[11] [S. R. Young, D. C. Rose, T. P. Karnowski, S.-H. Lim, and dimensions for vision architectures,” in International
R. M. Patton, “Optimizing deep learning hyper-parameters conference on machine learning. PMLR, 2013, pp. 115–
through an evolutionary algorithm. In: MLHPC’15 123.
Proceedings of the Workshop on Machine Learning in [27] F.-A. Fortin, F.-M. De Rainville, M.-A. Gardner, M.
High-Performance Computing Environments, Article No. Parizeau, and C. Gagn´e, “DEAP: Evolutionary Algorithms
4,” Oak Ridge National Laboratory, Oak Ridge Leadership Made Easy,” Journal of Machine Learning Research, vol.
Computing Facility (OLCF), Tech. Rep., 2015. 13, pp. 2171–2175, jul 2012.
[12] E. Real, S. Moore, A. Selle, S. Saxena, Y. L. Suematsu, J. [28] Ali, A., Lin, M., Andargie, M. S., & Azar, E. (2022).
Tan, Q. Le, and A. Kurakin, “Large-scale evolution of Demographics as Determinants of Building Occupants’
image classifiers,” arXiv preprint arXiv:1703.01041, 2017. Indoor Environmental Perceptions: Insights from a
[13] X. Xiao, M. Yan, S. Basodi, C. Ji, and Y. Pan, “Efficient Machine Learning Incremental Modeling and Analysis
Hyperparameter Optimization in Deep Learning Using a Approach. Journal of Computing in Civil
Variable Length Genetic Algorithm,” arXiv preprint Engineering, 36(4), 04022017.
arXiv:2006.12703, 2020. [29] Lin, M., Ali, A., Andargie, M. S., & Azar, E. (2021).
[14] J.-T. Tsai, J.-H. Chou, and T.-K. Liu, “Tuning the structure Multidomain drivers of occupant comfort, productivity, and
and parameters of a neural network by using hybrid well-being in buildings: Insights from an exploratory and
Taguchi-genetic algorithm,” IEEE Transactions on Neural explanatory analysis. Journal of Management in
Networks, vol. 17, no. 1, pp. 69–80, 2006. Engineering, 37(4), 04021020
[15] P.-W. Chen, J.-Y. Wang, and H.-M. Lee, “Model selection
of SVMs using GA approach,” in 2004 IEEE International
Joint Conference on Neural Networks (IEEE Cat. No.
04CH37541), vol. 3. IEEE, 2004, pp. 2035–2040.

0389
Authorized licensed use limited to: Arizona State University. Downloaded on April 09,2025 at 03:23:06 UTC from IEEE Xplore. Restrictions apply.