Page 35 - IJAMD-1-1

International Journal of AI
for Material and Design ML for quality improvement in L-PBF

products falls within the domain of a classification task. most representative features while eliminating irrelevant
Linear regression (LR) is a prominent regression model in and noisy information in the data. While unsupervised
supervised learning. Numerous ML models demonstrate learning reduces the labeling workload, it introduces
proficiency in handling classification tasks, such as neural challenges in precisely evaluating the model. In addition,
network (NN), support vector machine (SVM), decision the outcomes of unsupervised learning methods can be
tree (DT), and K-nearest neighbor (KNN). susceptible to the impact of arbitrarily selected initial values.

2.1.2. Unsupervised learning 2.1.3. Semi-supervised learning
In comparison to supervised learning, unsupervised Acquiring large, labeled datasets proves to be both
learning greatly alleviates the challenges associated costly and tedious in the context of supervised learning.
with the time-consuming and labor-intensive labeling Simultaneously, unsupervised learning heavily relies
process, as it does not require data labels for model on initial assumptions and poses challenges in terms of
training. Unsupervised learning broadly falls into two evaluation. Semi-supervised learning bridges these gaps by
main categories: clustering and dimensionality reduction. utilizing a combined approach that leverages both labeled
Representation clustering methods include K-means, and unlabeled data. In this way, labeled data serves as
hierarchical clustering (HC), Gaussian mixtures (GM), restrictions or introductions to the model, while unlabeled
among others. The input for clustering methods is the data prevents overfitting and reduces the laborious labeling
training data, and the output is a set of clusters. The process. The majority of semi-supervised methods are based
allocation of data points relies on a similarity assessment, on the smoothness assumption, low-density assumption,
21
which is crucial for clustering methods. For example, the and manifold assumption. The smoothness assumption
K-means algorithm, as illustrated in Figure 4, starts with illustrates that if any two data points are approximate in
randomly selected clustering center points and ends when the dataset, they will share the same label. The low-density
either the clusters remain unchanged or the maximum assumption serves as the low-dimensional counterpart of the
training iteration number is reached. Dimensionality smoothness assumption, asserting that the decision boundary
reduction represents mapping high-dimensional original line will not intersect the high-density area of data points. In
data into a low-dimensional subspace by capturing the manifold assumption, manifolds stand for topological
essential information in data, which contributes to easier spaces with local Euclidean dimensions, suggesting that
computation and visualization. Principal component the high-dimensional data space consists of multiple low-
analysis is a typical dimensionality reduction algorithm. It dimension manifolds. This assumption states that data
identifies projections between the input high-dimensional points within the same manifold share identical labels. Semi-
data and the output low-dimensional data that capture the supervised learning can be classified into two categories:
inductive methods and transductive methods. Inductive
methods aim to generate a classification model from the
input data. For example, a classifier is developed using labeled
data, which is then utilized to label the unlabeled data. On
the contrary, transductive methods focus on establishing
direct connections between data points without constructing
a model, lacking a clear phase of training and testing.
2.1.4. RL
In the realm of ML, RL is a dependent category that draws
inspiration from the human learning pattern of structuring
newly obtained knowledge through interactions with the
environment. Different from supervised and unsupervised
learning, RL involves a model that is not supervised;
instead, it obtains valuable numerical feedback in the form
of rewards. Throughout the training process, the RL model
constructs interactions aiming at gaining more rewards.
The field of RL encompasses two main components and
three core concepts. The two components involve the
interacting counterparts: the agent (the model) and the
Figure 4. The pseudocode of the K-means algorithm. environment, while the three concepts encompass the state

Volume 1 Issue 1 (2024) 29 https://doi.org/10.36922/ijamd.2301