Research Square (Research Square),
Год журнала:
2023,
Номер
unknown
Опубликована: Июнь 23, 2023
Abstract
Advancements
in
computing
and
storage
technologies
have
significantly
contributed
to
the
adoption
of
deep
learning
(DL)-based
models
among
machinelearning
(ML)
experts.
Although
a
generic
model
can
be
used
search
fora
near-optimal
solution
any
problem
domain,
what
makes
these
DL
modelscontext-sensitive
is
combination
training
data
hyperparameters.
Due
lack
inherent
explainability
HyperparameterOptimization
(HPO)
or
tuning
specific
each
art,science,
experience.
In
this
article,
we
explored
various
existing
methods
ways
identify
optimal
set
values
for
hyperparameters
specificto
along
with
techniques
realize
those
real-lifesituations.
The
article
also
includes
detailed
comparative
study
variousstate-of-the-art
HPO
using
Keras
Tuner
toolkit
highlights
observations
describing
how
performance
improvedby
applying
techniques.
Bioengineering,
Год журнала:
2024,
Номер
11(1), С. 77 - 77
Опубликована: Янв. 13, 2024
Accurate
classification
of
electromyographic
(EMG)
signals
is
vital
in
biomedical
applications.
This
study
evaluates
different
architectures
recurrent
neural
networks
for
the
EMG
associated
with
five
movements
right
upper
extremity.
A
Butterworth
filter
was
implemented
signal
preprocessing,
followed
by
segmentation
into
250
ms
windows,
an
overlap
190
ms.
The
resulting
dataset
divided
training,
validation,
and
testing
subsets.
Grey
Wolf
Optimization
algorithm
applied
to
gated
unit
(GRU),
long
short-term
memory
(LSTM)
architectures,
bidirectional
networks.
In
parallel,
a
performance
comparison
support
vector
machines
(SVMs)
performed.
results
obtained
first
experimental
phase
revealed
that
all
RNN
evaluated
reached
100%
accuracy,
standing
above
93%
achieved
SVM.
Regarding
speed,
LSTM
ranked
as
fastest
architecture,
recording
time
0.12
ms,
GRU
0.134
Bidirectional
showed
response
0.2
while
SVM
had
longest
at
2.7
second
phase,
slight
decrease
accuracy
models
observed,
98.46%
LSTM,
96.38%
GRU,
97.63%
network.
findings
this
highlight
effectiveness
speed
task.
Computers, materials & continua/Computers, materials & continua (Print),
Год журнала:
2024,
Номер
79(1), С. 19 - 46
Опубликована: Янв. 1, 2024
Hyperspectral
image
classification
stands
as
a
pivotal
task
within
the
field
of
remote
sensing,
yet
achieving
highprecision
remains
significant
challenge.In
response
to
this
challenge,
Spectral
Convolutional
Neural
Network
model
based
on
Adaptive
Fick's
Law
Algorithm
(AFLA-SCNN)
is
proposed.The
(AFLA)
constitutes
novel
metaheuristic
algorithm
introduced
herein,
encompassing
three
new
strategies:
weight
factor,
Gaussian
mutation,
and
probability
update
policy.With
adaptive
can
adjust
weights
according
change
in
number
iterations
improve
performance
algorithm.Gaussian
mutation
helps
avoid
falling
into
local
optimal
solutions
improves
searchability
algorithm.The
strategy
exploitability
adaptability
algorithm.Within
AFLA-SCNN
model,
AFLA
employed
optimize
two
hyperparameters
SCNN
namely,
"numEpochs"
"miniBatchSize",
attain
their
values.AFLA's
initially
validated
across
28
functions
10D,
30D,
50D
for
CEC2013
29
CEC2017.Experimental
results
indicate
AFLA's
marked
superiority
over
nine
other
prominent
optimization
algorithms.Subsequently,
was
compared
with
(FLA-SCNN),
Harris
Hawks
Optimization
(HHO-SCNN),
Differential
Evolution
(DE-SCNN),
(SCNN)
Support
Vector
Machines
(SVM)
using
Indian
Pines
dataset
Pavia
University
dataset.The
experimental
show
that
outperforms
models
terms
Accuracy,
Precision,
Recall,
F1-score
University.Among
them,
Accuracy
reached
99.875%,
98.022%.In
conclusion,
our
proposed
deemed
significantly
enhance
precision
hyperspectral
classification.
Atmosphere,
Год журнала:
2025,
Номер
16(4), С. 419 - 419
Опубликована: Апрель 4, 2025
Accurate
drought
prediction
is
crucial
for
optimizing
water
resource
allocation,
safeguarding
agricultural
productivity,
and
maintaining
ecosystem
stability.
This
study
develops
a
methodological
framework
short-term
forecasting
using
SPEI
time
series
(1979–2020)
evaluates
three
predictive
models:
(1)
baseline
XGBoost
model
(XGBoost1),
(2)
feature-optimized
variant
incorporating
Pearson
correlation
analysis
(XGBoost2),
(3)
an
enhanced
CPSO-XGBoost
integrating
hybrid
particle
swarm
optimization
with
dual
mechanisms
of
binary
feature
selection
parameter
tuning.
Key
findings
reveal
spatiotemporal
patterns:
temporal-scale
dependencies
show
all
models
exhibit
limited
capability
at
SPEI-1
(R2:
0.32–0.41,
RMSE:
0.68–0.79)
but
achieve
progressive
accuracy
improvement,
peaking
SPEI-12
where
attains
optimal
performance
0.85–0.90,
0.33–0.43)
18.7–23.4%
error
reduction
versus
baselines.
Regionally,
humid
zones
(South
China/Central-Southern)
demonstrate
peak
(R2
≈
0.90,
RMSE
<
0.35),
while
arid
regions
(Northwest
Desert/Qinghai-Tibet
Plateau)
dramatic
improvement
from
0.35,
>
1.0)
to
0.85,
52%).
Multivariate
probability
density
confirms
the
model’s
robustness
through
capture
nonlinear
atmospheric-land
interactions
reduced
parameterization
uncertainties
via
intelligence
optimization.
The
CPSO-XGBoost’s
superiority
stems
synergistic
optimization:
enhances
input
relevance
adaptive
tuning
improves
computational
efficiency,
collectively
addressing
climate
variability
challenges
across
diverse
terrains.
These
establish
advanced
early
warning
systems,
providing
critical
support
climate-resilient
management
risk
mitigation
spatiotemporally
predictions.