Page 157 - Intelligent Digital Oil And Gas Fields

120 Intelligent Digital Oil and Gas Fields

Table 4.4 Summary of the Most Popular Methods and Classifiers in ML, as
Representative of the Three Main Paradigms Described Previously, With Some
Suggested References for Further Reading—cont’d
ML Family
Type of Suggested
ML Paradigm Problem Technique Reading

Unsupervised Clustering • K-means James et al.
learning • Hierarchical (2014)
Hastie et al.
(2011)
Leskovec
et al. (2014)
Dimension • Principal component analysis Tabachnick
reduction (PCA) and Fidell
• Factor analysis (2013)
• Multidimensional scaling
(MDS)
Reinforcement • Markov decision process Sutton and
learning (MDP) Barto
(1998)

representative elements, called centroids. With dimensionality reduction
methods on the other hand, the analyst is aiming to represent the com-
plex numerical model with a reduced or compressed set of (principal)
components (usually referred to as eigenvalues and eigenvectors), which
can still adequately represent the observation domain x 1 , …, x n , while
significantly reducing the computational effort and complexity. The
clustering and dimensionality reduction methods often fall under the cat-
egory of MVA methods (Tabachnick and Fidell, 2013).
• RL is a paradigm where the information available in the training of the
ML model can be viewed as a cross-section between supervised and
unsupervised learning. The RL methods usually leverage the ideas
and algorithms from the area of control theory (e.g., optimal, robust con-
trol, closed-loop control, and variance reduction). The mathematical
foundation of ML methods represents Markov decision processes
(MDP), similar to Markov chains and its Monte Carlo approximations.
The complexity of these techniques goes beyond the scope of this book;
to learn more, see Sutton and Barto (1998).