A Primer on Machine Learning Applications in Civil Engineering

Paresh Chandra Deka

• Publisher: CRC Press
• Year: 2019
• ISBN:
  9780429836657
• Edition: 1
• Topic: Informatik
• Subtopic: Informatik Allgemein

Machine learning has undergone rapid growth in diversification and practicality, and the repertoire of techniques has evolved and expanded. The aim of this book is to provide a broad overview of the available machine-learning techniques that can be utilized for solving civil engineering problems. The fundamentals of both theoretical and practical aspects are discussed in the domains of water resources/hydrological modeling, geotechnical engineering, construction engineering and management, and coastal/marine engineering. Complex civil engineering problems such as drought forecasting, river flow forecasting, modeling evaporation, estimation of dew point temperature, modeling compressive strength of concrete, ground water level forecasting, and significant wave height forecasting are also included.

Features

• 
  
  

• Exclusive information on machine learning and data analytics applications with respect to civil engineering

• 
  
  

• Includes many machine learning techniques in numerous civil engineering disciplines

• 
  
  

• Provides ideas on how and where to apply machine learning techniques for problem solving

• 
  
  

• Covers water resources and hydrological modeling, geotechnical engineering, construction engineering and management, coastal and marine engineering, and geographical information systems
• Includes MATLAB® exercises

1

Introduction

1.1 Machine Learning

Machine learning is related to artificial learning and has been getting tremendous attention from researchers across the globe since its inception. It is indeed a very powerful approach to data analysis and modeling for various applications. The fundamental feature of machine learning algorithms is that they can learn from empirical data and can be used where modeled phenomena are hidden, non-evident, or not very well explained. There are many different algorithms available in machine learning using many methods of nonparametric statistics, artificial intelligence research, and computer science. Here, the most commonly used MLAs for Civil Engineering related problems are Artificial Neural Network (ANN), Fuzzy Logic (FL), Genetic Algorithm (GA), and Support Vector Machine (SVM).

In the context of machine learning, ANN, FL, GA, and SVM are important models. These are not only approaches to develop artificial intelligence but are also used as universal nonlinear and adaptive tools for solving data-driven classification and regression problems. In this regard, machine learning is seen as an applied scientific discipline, while the general properties of statistical learning from data and the mathematical theory of generalization from experience are more relevant.

The complexity necessary to mathematically model real-world problems has compelled human civilization to search for nature-inspired computing tools. The evolution of such computing tools revolves around the information processing characteristics of biological systems. In contrast to conventional computing, these tools are rather ‘soft’ as they lack exactness and are, therefore, placed under the umbrella of a multidisciplinary field called soft computing.

Soft computing is an emerging collection of methodologies which aims to exploit tolerance for imprecision, uncertainty, and partial truth to achieve robustness, tractability, and cost savings. It is a branch of computational intelligence research that employs a variety of statistical, probabilistic, and optimization tools to learn from past examples and to then use that prior training to classify new data, identify new patterns, or predict novel trends. Soft computing tools exploit the reasoning, intuition, consciousness, wisdom, and adaptability to changing environments possessed by human beings for developing computing paradigms like FL, ANNs, GAs, and SVMs. The integration of these techniques into the computing environment has given impetus to the development of smarter and wiser machines possessing logical and intuitive information processing capabilities equivalent to human beings. These techniques, whether complementing each other or working on their own, are able to model complex or unknown relationships which are either nonlinear or noisy. Soft computing techniques have a self-adapting characteristic, paving the way for the development of automated design systems. A synergistic partnership exploiting the strengths of these individual techniques can be harnessed for developing hybrid computing tools.

Among the forerunners in the field of soft computing is the Artificial Neural Network (ANN). Inspired by the functioning of a human brain, they have immense potential in modeling functional relationships which are either too complex or unknown in nature. The ANN model is a system of interconnected computational neurons arranged in an organized fashion to carry out the extensive computing necessary to perform the mathematical mapping. Unlike conventional methods of computing which are based on predefined rules, ANNs work on vague functional relationships through a process of learning from experience and examples.

Fuzzy Logic (FL) was conceptualized by Zadeh in the year 1965. It was inspired by how human beings make decisions when dealing with knowledge that is inexact, imprecise, and vague in nature; FL in a way emulates human expertise in solving a particular problem. Genetic Algorithms (GAs) are stochastic search and optimization tools which aim to find the optimal solution to a problem which has many sub-optimal solutions. They require little information about the problem to be solved and can effectively work with complex constraints and discrete variables. GAs working on the operators of natural evolution viz., reproduction, crossover, and mutation were inspired by Darwin’s principle of ‘survival of the fittest’ and are able to find the global optimal solution to a particular problem.

In recent years, a novel machine learning technique called Support Vector Machine (SVM), proposed by Vapnik and based on statistical learning theory, has been successfully applied to pattern classification and regression estimation problems. Initially, SVMs were developed for solving classification problems and later were expanded by Vapnik to regression problems. Unlike ANNs, which are based on the Empirical Risk Minimization (ERM) induction principle, SVM implements the Structural Risk Minimization (SRM) principle. It is well known that ANN has the problem of overfitting and that the solution may get trapped in local minima, whereas the SRM principle aims to minimize an upper bound on the generalization error; and thus SVM will have higher prediction capabilities on unseen data sets. Further, SVM formulation will lead to the solution of a quadratic optimization problem with linear inequality constraints and the problem will have a global optimal solution. Combined with high generalization ability, SVM becomes a very attractive method.

This book is oriented to the approach of data-driven modeling to both spatial and temporal field data. A small component of the book related to fundamentals is devoted to machine learning algorithms. The main focus is on various models that have been utilized by the authors in various domains of Civil Engineering. This book is targeted to graduate and doctoral students of Civil Engineering, as well as Earth sciences departments and researchers interested in machine learning methods and applications.

The objective of machine learning (ML) is to develop algorithms which allow computers to learn. This learning is based on experience gained and abilities to generalize previous scenarios into new conditions. Learning abilities are important for human intelligence. In artificial intelligence, the main challenge is to equip the machine with this ability, either by implementing a set of algorithms or by a stand-alone robot.

In the early days of ML, the idea was to use it to make machines adaptable, interactive, and able to learn from experience. ML continues to bring many challenging problems to fundamental research. Speech recognition, computer vision, and feature extraction are the new scientific branches where ML is essential. Computational learning, which studies the properties of learning from the experimental/empirical/field data, is also becoming an important research field.

The importance of ML may be attributed to scientific research changes over the past few decades. The gathering of relevant data in the world around us has drastically improved with the support of recent technological advances, and Civil Engineering is a field which immensely benefits from these advances. Various sensors are put into systems to measure countless parameters, and they can be organized in wireless networks to provide huge volumes of information in real time.

The storage of data is a technical and engineering problem, whereas the understanding of the underlying phenomena is a scientific challenge. Recently, scientific research has been inclined towards a data-driven approach and this data-driven modeling has to respond to the problem accordingly, with the help of ML.

1.2 Learning from Data

At the beginning of learning, developing a system or an algorithm which can learn and generalize from the data is the formulation of an appropriate mathematical framework. In the majority of cases, an observation can be presented as a pair of entities—one is input space and the other is output space. Hence, empirical knowledge can be formulated as a set of these input–output pairs. Both input and output data can be encoded as multidimensional vectors.

From the observation and collected data, it can be assumed that some kind of underlying phenomena links inputs to outputs. This dependence is assumed as D, such that D maps input x to y. A deterministic mapping is a natural way to link vector spaces, although it is not realistic to restrict real-world processes which generate the data to be ideally deterministic. As there are many factors influencing data and measurement processes in the real world, the settings become stochastic. Hence, the role of probabilistic distribution P(x,y) is inevitable for representation of the generated data in order to achieve an acceptable description of the process. Usually, the explicit form of this type of distribution is unknown. However, some kind of inference from the available data set (x,y) generated by P(x,y) is to make sure that the data set is consistent and properly represented to provide reliable knowledge about P(x,y). Hence, it can be assumed that (x,y) are independent and identically distributed data sampled from the same population.

Machine learning developed algorithms are able to predict the outputs for previously unknown inputs without making restrictive assumptions about P(x,y). Although, whereas some ideas are purely algorithmic and distribution independent, it is essential to have an empirical data set that is representative of the underlying processes. This means the new samples are from the very same distribution/population for prediction.

1.3 Research in Machine Learning: Recent Progress

The field of engineering is a creative one. The problems encountered in this field are generally unstructured and imprecise, influenced by intuitions and past experiences of ...