A Machine Learning-based system for berth scheduling at bulk terminals

Highlights

• Novel machine learning-based system for supporting berthing operations in bulk ports.

• The system recommends the best algorithm in most of the studied cases.

• Different meta-features for the berth allocation problem are investigated.

• The system exhibits a relevant robustness when tackling new and unfamiliar scenarios.

Abstract

The increasing volume of maritime freight is presented as a challenge to those skilled terminal managers seeking to maintain or increase their market share. In this context, an efficient management of scarce resources as berths arises as a reasonable option for reducing costs while enhancing the productivity of the overall terminal. In this work, we tackle the berth scheduling operations by considering the Bulk Berth Allocation Problem (Bulk-BAP). This problem, for a given yard layout and location of the cargo facilities, aims to coordinate the berthing and yard activities for giving service to those vessels arriving at the terminal. Considering the multitude of scenarios arising in this environment and theNo Free Lunch theorem, the drawback concerning the selection of the best algorithm for solving the Bulk-BAP in each particular case is addressed by a Machine Learning-based system. It provides, based on the scenario at hand, a ranking of algorithms sorted by appropriateness. The computational study shows an increase in the quality of the provided solutions when the algorithm to be used is selected according to the features of the instance instead of selecting the best algorithm on average.

Introduction

Maritime freight remains as the most important transportation mean for worldwide commerce, estimated at over 80% of world trade. A report by the European Union (EU) published in 2011 (see UE, 2011) indicated that the 37.8% of the transport of goods within the EU is made through the sea. In this context, according to a report from the United Nations Conference on Trade and Development (UNCTAD) presented in 2015 (UNCTAD, 2015), the world economy is currently passing through a slow recovery from the economic crisis. Moreover, in 2016, the global seaborne shipments have increased 2.1% and the world fleet grew by 3.5% (in terms of dead-weight tons). Shipping is the backbone of international trade and given the fact that some countries and regions are in a slow recovery, it is necessary to design and implement systems for decision support to maximize the benefits by optimizing the use of available resources. The optimization of these resources will occur through the minimization of the costs associated with the vessels handling operations. Those costs are classified by (Stopford, 2003) into five major groups:

• Operating costs: these costs are related to daily operations of the vessels independent from the business domain, e.g. crew cost, general stores, navigation and comms service, communications, etc.

• Periodic maintenance costs: is a provision set aside to cover the cost of interim dry-docking and special surveys, e.g. port charges, tugs, agency, all steel replacement, piping and valves, etc.

• Transportation costs: this type of costs is dependent on the use of the vessel, for example, port taxes or fuel consumption.

• Capital costs: they are related to the way in which the purchase or use of the vessel was financed.

• Costs associated with cargo handling: represent those costs related to cargo operations, loading and unloading of the goods.

Considering the above, it can be remarked that a relevant part of the total costs heavily relies on the provided services. Thus, a poor planning due to an underutilization of resources has a direct impact on the overall benefits of the port leading to a loss of competitiveness. In this context, an inefficient utilization of key resources such as berths leads to longer service times while reducing the productivity of the terminal. As indicated in the related literature (Lalla-Ruiz, 2016), the decisions related to the management of berths have a direct impact on the rest of the problems taking place at the terminal. The above entails the need to provide suitable and optimized solutions in a reasonable time frame.

In the related literature, the Berth Allocation Problem (BAP) consists of assigning positions and berthing times to arriving vessels at the port with the aim of optimizing a given objective function. It has been widely studied in the literature (Bierwirth & Meisel, 2015) presenting different variants according to spatial and time constraints. Within the spatial restrictions, the quay can be considered (i) discrete: the quay is divided into equal parts called berths (Cordeau, Laporte, Legato, Moccia, 2005, Hansen, Oguz, Mladenovic, 2008, Lalla-Ruiz, Melián-Batista, Moreno-Vega, 2012a, Monaco, Sammarra, 2007); (ii) continuous: the quay is treated as a continuous section allowing vessels to berth at any point within it (Frojan, Correcher, Álvarez-Valdes, Koulouris, Tamarit, 2015, Lee, Chen, Cao, 2010); and (iii) hybrid: the quay is divided into sections enabling a vessel to occupy more than one section (Cordeau, Laporte, Legato, Moccia, 2005, Umang, Bierlaire, Vacca, 2013). Moreover, concerning the temporal restrictions, the BAP can be treated as (iv) static: all the vessels are at the terminal when the planning is going to be conducted (Imai, Nagaiwa, & Tat, 1997); (v) dynamic: the vessels arrive along the planning horizon (Cordeau, Laporte, Legato, Moccia, 2005, Imai, Nishimura, Papadimitriou, 2001); and (vi) time-dependent: the availability of the berths changes along the time horizon (Lalla-Ruiz, Expósito-Izquierdo, Melián-Batista, Moreno-Vega, 2016, Xu, Li, Leung, 2012). Furthermore, in the BAP the arrival time has also been considered as a decision variable to maximize the reliability of the schedule minimizing vessel delayed departures (Golias, Saharidis, Boile, Theofanis, & Ierapetritou, 2009). Moreover, handling and arrival times are considered stochastic parameters to minimize total waiting time of calling vessels (Zhou & Kang, 2008) or to maximize berth efficiency and maintain the integrity of the schedule (Golias, 2011). Among the previously cited papers, Umang et al. (Umang, Bierlaire, & Vacca, 2013) address the Berth Allocation Problem in bulk ports (Bulk-BAP) which considers a hybrid quay layout that divides the quay into sections and where each section can only be occupied by a vessel at each instant of time, but a vessel can occupy more than one section. Regarding temporal constraints, the Bulk-BAP is included in the dynamic variant category.

Considering the high dynamism existing in maritime terminals, especially in some operations at the seaside, such as the allocation of berths, the operators have to deal with multiple and diverse scenarios whose solution implies selecting the best algorithm from an algorithmic bed or using different strategies depending on their respective characteristics. The aforementioned issue, from the viewpoint of decision support systems, requires not only having efficient algorithms but also machine learning mechanisms for selecting the most appropriate algorithm adjusted to the conditions of the scenario at hand in such a way that all the involved resources are properly managed. In order to address that, Meta-Learning (Brazdil, Carrier, Soares, Vilalta, 2008, Pappa, Ochoa, Hyde, Freitas, Woodward, Swan, 2014), a sub-field of Machine Learning, investigates and proposes different strategies to determine which algorithm is better suited for being use according to a given problem instance. This way, the decision support system is enabled to maximize the benefits of its algorithms by selecting the most suitable one to solve a given scenario based on its corresponding casuistry.

In view of the previous discussion, the contributions of this paper are summarized in the following points:

• A problem instance generator for the Bulk-BAP that allows defining problem instances for different possible scenarios is proposed, e.g. congested scenarios. This generator simplifies the generation of instances sets by performing all the possible combinations of values that the parameters composing the instances can take.

• In order to select an appropriate solution approach adjusted to the problem instance at hand, a Machine Learning-based system for the Bulk-BAP is proposed. It is able to provide a ranking of algorithms from a pre-established pool of algorithms based on the instance to be solved. That is to say, the choice of the algorithm is made on the basis of the characteristics of the instance and not on the basis of the average performance of the algorithms. The computational results report relevant advantages when using this system instead of the best algorithm on average, supporting the possibility of applying this type of approaches to real-world environments.

The organization of the paper is described below. Section 2 describes the Bulk Berth Allocation Problem as well as the generation of the benchmark suite used in this work. In Section 3 the Algorithm Selection Problem is detailed. Section 4 describes the main features of the Machine Learning-based system and the algorithms used to evaluate the system. The results obtained by our proposed system are discussed in Section 5. Finally, Section 6 presents the conclusions together with future research lines.