Machine learning and data analytics can be used to inform technical, commercial and financial decisions in the maritime industry. Applications of Machine Learning and Data Analytics Models in Maritime Transportation explores the fundamental principles of analysing maritime transportation related practical problems using data-driven models, with a particular focus on machine learning and operations research models.
Data-enabled methodologies, technologies, and applications in maritime transportation are clearly and concisely explained, and case studies of typical maritime challenges and solutions are also included. The authors begin with an introduction to maritime transportation, followed by chapters providing an overview of ship inspection by port state control, and the principles of data driven models. Further chapters cover linear regression models, Bayesian networks, support vector machines, artificial neural networks, tree-based models, association rule learning, cluster analysis, classic and emerging approaches to solving practical problems in maritime transport, incorporating shipping domain knowledge into data-driven models, explanation of black-box machine learning models in maritime transport, linear optimization, advanced linear optimization, and integer optimization. A concluding chapter provides an overview of coverage and explores future possibilities in the field.
The book will be especially useful to researchers and professionals with expertise in maritime research who wish to learn how to apply data analytics and machine learning to their fields.
Other keywords: ships; regression analysis; sea ports; maritime transportation; machine learning; pattern classification; decision support systems; linear programming; marine safety; unsupervised learning; data analytics model
Maritime transportation is the transport of passengers and cargoes by sea which can be dated back to ancient Egypt times. It is the backbone of global trade, manufacturing supply chain, and world economy, as about 80% of world trade by volume is carried out by ocean-going vessels [1]. Even during the ever-hard time brought about by the COVID-19 pandemic, the maritime transport still plays a key role in moving various goods and products, especially the necessary living and medical supplies, around the world. The world economic development has a critical impact on maritime transport. According to the review of maritime transport produced by the United Nations Conference on Trade and Development (UNCTAD), the international maritime trade volumes from 1970 to 2020 are shown in Table 1.1 [1].
Maritime transport is a highly globalized industry in terms of operation and management. For ship operation, ocean-going vessels sail on the high seas from the origin port in one country/region to the destination port in another country/region. For ship management, parties responsible for ship ownership, crewing, and operating may locate in different countries and regions. Even the country of registration, i.e., ship flag state, may not have a direct link and connection with a ship's activities as the ship may not frequently visit the ports belonging to its flag state. For inland countries such as Mongolia, the ships registered under it never visit its ports. Such complex and disintegrated nature of the shipping industry makes it hard to control and regulate international shipping activities, and thus pose danger to maritime safety, the marine environment, and the crew and cargoes carried by ocean-going vessels.
This chapter aims to clarify the basic issues of data-driven modeling and its application in maritime transport. A predictive problem is first introduced, and the predictive analysis in the field of maritime transport is then discussed with typical examples given. A comparison between the classic method and data-driven modeling to deal with prediction tasks is also provided. Finally, typical data-driven modeling approaches are briefly discussed.
This chapter aims to thoroughly introduce and discuss the major issues of data-driven models, with a focus on machine learning (ML) models and their construction procedure. We first compare three popular data-driven models, namely, statistical model, ML model, and deep learning (DL) model. Then, the whole procedure of developing a data-driven model will be provided. Key elements from various aspects of the whole process will be covered in detail.
Linear regression aims to learn a linear model that can predict the target using the features as accurately as possible. The assumption of linear regression models is that the target is linearly correlated with the features, i.e., the regression function E(y|x) is linear in x, where E(⋅) denotes expectation. If the assumption can (almost) be satisfied, linear regression can be comparable or can even outperform fancier non-linear models. The linear regression model is one of the most classic models for prediction tasks, and it is still widely used in the computer and big data era, thanks to its intuitiveness and interpretability in particular. In the following sections, we first introduce simple linear regression models (with a single feature) and the least squares method, which aims to find the optimal parameters of a linear regression model by minimizing the sum of squares of the residuals. Then, we discuss multiple linear regression (with more than one feature) and its extension. Finally, we introduce shrinkage linear regression models.
This chapter introduces the basics of Bayesian network (BN) classifiers that are used to address classification problems. Naive Bayes classifier is first presented, where a simplified (but unrealistic) assumption that the features are conditionally independent and are of equal importance is made. To weaken the assumption so as to improve the classification accuracy, semi-naive Bayes classifiers are then presented, where part of the dependencies between the features is considered. Finally, BN in more general form is introduced.
This chapter first introduces one of the most popular machine learning models for classification tasks called support vector machine (SVM). Then, kernel trick to improve its prediction accuracy while reducing the computation burden is discussed. The extension of SVM to address regression tasks, which is called support vector regression, is then presented.
This chapter aims to introduce a widely-used machine learning model for both regression and classification tasks called artificial neural network (ANN). Basic concepts of ANNs are first covered, and then typical algorithms and tricks for ANN training are introduced. Finally, deep learning models, as an important type of neural network, are briefly discussed.
This chapter aims to introduce several machine learning models based on a tree structure called the decision tree, which are widely believed to be among the most popular methods for both classification and regression tasks. The basic structure and concepts as well as the tree-growing algorithms of a single decision tree are first introduced. As a single decision tree is prone to over-fitting, ensemble models consisting of a certain number of decision trees are developed. Random forest-based on bagging and gradient boosting decision trees based on boosting will be introduced as the representatives of ensemble models on decision trees.
Association rule learning is a rule-based and unsupervised machine learning method, which is widely used to discover interesting relations between items (i.e., records) in a database and also to explore how and why these items are connected. A widely known application of association rule learning is to analyze the purchased items in market basket transactions, which aims to identify what goods are often bought together in one transaction so as to adjust sales strategies to increase sales. For example, a famous story is that on Friday nights, the sales of diapers and beer were correlated in Walmart: a bottle of beer was often bought when diapers were bought. The explanation was that working men were asked to pick up diapers on their way home from work, and they would also buy a bottle of beer for themselves at the same time. Based on this finding, Walmart put the shelves of these two goods close to each other on Friday nights, and the sales volume of both increased greatly. In the above example, diapers and beer are correlated in the following way: diapers → beer, meaning that if diapers are bought, beer is highly like to be bought in the same transaction. This is a basic form of association rule we are going to cover in this section. Actually, association rule learning is widely used to mine the relations of items from transaction databases, and the rules generated are used to guide the activities of personalized product recommendation, shelf placement, combined coupon dispatching, and bundle sales in the retailing industry.
Cluster analysis, or clustering, is a general task aiming to group a given set of examples into several groups (i.e., clusters) following given criteria, such that the examples in the same group are as close as to each other, and the examples in different groups are as different as from each other. Generally, clustering works in the context of unsupervised learning, where only the features of the examples are known while there is no target defined or targets are unknown. It aims to divide the whole data set into several mutually exclusive and complementary clusters, so as to mine the properties of the clusters formulated where such properties are represented by cluster labels.
This chapter aims to discuss classic and emerging approaches adopted in existing academic literature to address practical problems in maritime transport. First, widely studied practical problems in the maritime industry are summarized according to review papers. Then, classic methods that are widely adopted by the related studies are introduced. After that, data-driven methods, as a typical type of emerging approaches, are discussed. Especially, several examples of applying data-driven models to address prediction tasks in maritime studies are given. Finally, several issues regarding the application of data-driven models to address practical problems in maritime transport are presented with various examples in port state control (PSC) provided.
As mentioned in section 12.3.4 of chapter 12, one of the issues regarding the target of applying data-driven models to solve practical problems in maritime transportation is that the predicted target may not comply with shipping domain knowledge. The term "domain knowledge" refers to rules and common senses widely believed by the practitioners. Domain knowledge is based on the practitioners' understanding of the disciplines and activities of the industry, and is gained from longtime experience of the practitioners in this industry as well as their own expertise, professions, specializations, and judgment. For example, in the maritime industry, regarding the activity of ship selection for inspection by PSC (port state control), it is generally believed that given all other conditions being equal, an older ship would have a larger number of deficiencies than a younger ship. Regarding ship fuel consumption prediction, ship sailing speed is the most significant determinant, and it is widely believed that a ship's fuel consumption rate is proportional to its sailing speed to the power of α = 3, i.e., r ∝ β × v α, where r is the hourly or daily fuel consumption, β is a coefficient, and v is the average sailing speed. In practice, α can be higher than 3, especially for large vessels like container ships where it can be 4, 5, or even higher. In the former example, if there is a ship risk prediction model that gives opposite prediction results, i.e., a younger ship has more deficiencies than an older ship under the condition that the other features of these two ships are identical, the prediction results as well as the prediction model are expected to be hardly accepted or used by the port state officers, because they may conclude that the proposed model is inaccurate and unfair based on such prediction. Similarly, in the latter example, if the predicted ship fuel consumption rate does not take such convex and increasing relationship with ship sailing speed, it may not convince the users. Given the condition that practitioners in the conservative while classic maritime industry might be reluctant to replace the current rule-based decision support systems with data-driven ones, it is thus of vital importance to make sure that the prediction results given by the data-driven models to address practical problems comply with the corresponding shipping domain knowledge. Otherwise, the practitioners are more likely to be very skeptical of the models together with their results, and thus are not willing to use them to assist their decision-making.
One way to guarantee that the developed data-driven models constructed from practical data are in compliance with shipping domain knowledge is to explicitly impose the constraints when developing the models. The following two sections in this chapter introduce some initial thoughts on how to incorporate shipping domain knowledge into the development of data-driven models used to deal with a specific problem in maritime transportation. To be specific, section 13.1 discusses how to consider feature monotonicity into a tree-based model developed to predict ship risk, and section 13.2 discusses how to jointly consider feature monotonicity and convexity into ship fuel consumption rate prediction.
In addition to model fairness, another serious drawback of using ML (machine learning) models to address practical problems in maritime transport is related to the black-box property of most ML models: despite their success in many real-world applications to not only the maritime industry but also other industries including defense, medicine, finance, and law, thanks to their high prediction performance, they are opaque in terms of explainability, which makes users and even developers difficult to understand, trust, and manage such powerful AI (artificial intelligence) applications. Especially, when decisions derived from black-box systems based on ML models affect human's life, safety, and the environment, there is a more urgent need for explaining and understanding how such decisions are furnished by AI methods. Especially, in the relatively traditional and conservative maritime industry, decision-makers and stakeholders are more likely to be reticent to adopt decision support tools powered by new technologies such as AI which they can hardly interpret, control, and thus trust. This chapter aims to first introduce the necessity of explaining black-box ML models in the maritime industry, especially in the case of PSC (port state control). Then, popular methods to achieve black-box model explanation are introduced.
We will discuss linear optimization in this chapter, as linear optimization is important in itself and is the foundation of more advanced optimization techniques.
Basic linear optimization models have been introduced in Chapter 15. In this chapter, more advanced linear optimization models will be covered.
In linear optimization, we assume that all of the decision variables are continuous. However, in reality, some are not. For example, the number of crude oil tankers used to transport crude oil from the Middle East to the US is an integer, and rounding down 6.5 ships to 6 ships can lead to considerable errors. Therefore, a natural extension to linear optimization models is integer linear optimization models, which are the same as linear optimization models except that the decision variables can only take integer values. We also have mixed-integer linear optimization models, in which some decision variables can only take integer values, and the others are continuous. We often use "integer optimization models" to refer to integer linear optimization models or both integer linear optimization models and mixed-integer linear optimization models. However, one should keep in mind that integer linear optimization models are not linear; in other words, integer linear optimization models belong to the category of nonlinear optimization models.
We use ℤ+ to represent the set of non-negative integers. Hence, x ∈ ℤ + means that x is non-negative and can only take integer values.
Machine Learning and Data Analytics in Maritime Studies: Models, Algorithms, and Applications aims to explore the fundamental principle of analyzing practical problems in maritime transportation using data-driven models, especially using machine learning (ML) models and operations research models. The book first introduces the state-of-the-art data-enabled methodologies, technologies, and applications in maritime transportation in a way that is easy to understand by maritime researchers and practitioners. To achieve this, plain words are used to present the algorithms and models, and real examples of ship inspection by port state control and operations planning of container ships are accompanied. By doing so, readers are expected to learn how to solve practical problems in maritime transportation by data-driven models while taking shipping domain knowledge and black box model explanations into consideration.
Errata Sheet for Applications of Machine Learning and Data Analytics Models in Maritime Transportation'
There is an errata sheet available for 'Applications of Machine Learning and Data Analytics Models in Maritime Transportation'.