Tommaso Urli

Feature-based tuning of simulated annealing applied to the curriculum-based course timetabling problem

We consider the university course timetabling problem, which is one of the most studied problems in educational timetabling. In particular, we focus our attention on the formulation known as the curriculum-based course timetabling problem, which has been tackled by many researchers and for which there are many available benchmarks. The contribution of this paper is twofold. First, we propose an effective and robust single-stage simulated annealing method for solving the problem. Secondly, we design and apply an extensive and statistically-principled methodology for the parameter tuning procedure. The outcome of this analysis is a methodology for modeling the relationship between search method parameters and instance features that allows us to set the parameters for unseen instances on the basis of a simple inspection of the instance itself. Using this methodology, our algorithm, despite its apparent simplicity, has been able to achieve high quality results on a set of popular benchmarks. A final contribution of the paper is a novel set of real-world instances, which could be used as a benchmark for future comparison.

Single- and multi-objective genetic programming: new bounds for weighted order and majority

We consolidate the existing computational complexity analysis of genetic programming (GP) by bringing together sound theoretical proofs and empirical analysis. In particular, we address computational complexity issues arising when coupling algorithms using variable length representation, such as GP itself, with different bloat-control techniques. In order to accomplish this, we first introduce several novel upper bounds for two single- and multi-objective GP algorithms on the generalised Weighted ORDER and MAJORITY problems. To obtain these, we employ well-established computational complexity analysis techniques such as fitness-based partitions, and for the first time, additive and multiplicative drift. The bounds we identify depend on two measures, the maximum tree size and the maximum population size, that arise during the optimization run and that have a key relevance in determining the runtime of the studied GP algorithms. In order to understand the impact of these measures on a typical run, we study their magnitude experimentally, and we discuss the obtained findings.

Experimental supplements to the computational complexity analysis of genetic programming for problems modelling isolated program semantics

In this paper, we carry out experimental investigations that complement recent theoretical investigations on the runtime of simple genetic programming algorithms [3, 7]. Crucial measures in these theoretical analyses are the maximum tree size that is attained during the run of the algorithms as well as the population size when dealing with multi-objective models. We study those measures in detail by experimental investigations and analyze the runtime of the different algorithms in an experimental way.

Feature-based tuning of single-stage simulated annealing for examination timetabling

We propose a simulated annealing approach for the examination timetabling problem, as formulated in the 2nd International Timetabling Competition. We apply a single-stage procedure in which infeasible solutions are included in the search space and dealt with using suitable penalties. Upon our approach, we perform a statistically-principled experimental analysis, in order to understand the effect of parameter selection on the performance of our algorithm, and to devise a feature-based parameter tuning strategy, which can achieve better generalization on unseen instances with respect to a one-fits-all parameter setting. The outcome of this work is that this rather straightforward search method, if properly tuned, is able to compete with all state-of-the-art specialized solvers on the available instances. As a byproduct of this analysis, we propose and publish a new, larger set of (artificial) instances that could be used for tuning and also as a ground for future comparisons.

On the performance of different genetic programming approaches for the sorting problem

In genetic programming, the size of a solution is typically not specified in advance, and solutions of larger size may have a larger benefit. The flexibility often comes at the cost of the so-called bloat problem: individuals grow without providing additional benefit to the quality of solutions, and the additional elements can block the optimization process. Consequently, problems that are relatively easy to optimize cannot be handled by variable-length evolutionary algorithms. In this article, we analyze different single- and multiobjective algorithms on the sorting problem, a problem that typically lacks independent and additive fitness structures. We complement the theoretical results with comprehensive experiments to indicate the tightness of existing bounds, and to indicate bounds where theoretical results are missing.

Hybrid meta-heuristics for combinatorial optimization

Combinatorial optimization problems arise, in many forms, in various aspects of everyday life. Nowadays, a lot of services are driven by optimization algorithms, enabling us to make the best use of the available resources while guaranteeing a level of service. Examples of such services are public transportation, goods delivery, university time-tabling, and patient scheduling. The fields of meta-heuristics, artificial intelligence, and operations research, have been tackling many of these problems for years without much interaction. However, in the last few years, such communities have started looking at each other’s advancements, in order to develop optimization techniques that are faster, more robust, and easier to maintain. This effort gave birth to the fertile field of hybrid meta-heuristics. In this thesis, we analyze some of the most common hybrid meta-heuristics approaches, and show how these can be used to solve hard real-world combinatorial optimization problems, and what are the major advantages of using such techniques. This thesis is based on results obtained by working together with many local and international researchers, and published in a number of peer-reviewed papers. School: University of Udine Supervisors: Luca Di Gaspero Graduated: Friday, April 4, 2014 Link to full text: http://www.a4cp.org/sites/default/files/tommaso_urli_-_hybrid_metaheuristics_for_combinatorial_optimisation.pdf Constraints (2015) 20:473 DOI 10.1007/s10601-015-9209-7 * Tommaso Urli Tommaso.Urli@nicta.com.au 1 NICTA, Canberra, Australia

Vehicle routing problems with deliveries split over days

In the classical (Capacitated) Vehicle Routing Problem (CVRP) we seek to use a fleet of trucks to deliver goods to customers at minimal cost, and customers’ requests must be fulfilled in a single visit. In the Split Delivery Vehicle Routing Problem (SDVRP), we are allowed to visit a customer multiple times to fulfil their demand. In this paper, we study a variant of SDVRP where customers have to be served over a multi-period horizon, and each customer may require service on one or more days. We consider the possibility of splitting a request not only between vehicles on the same day but also across consecutive days. The objective is to find a set of routes for each day of the planning horizon so that the total travelling cost is minimised, and the total amount delivered meets customers’ demand. We show that this Multi-Day version of SDVRP is structurally different to the usual SDVRP. We also consider a Fleet Size and Mix variant of the problem, where a fixed cost for using a vehicle at all during the horizon is added to the objective function. We propose a Mixed Integer Linear Programming approach to solve the Multi-Day SDVRP, together with some valid inequalities to enhance it. Furthermore, we develop a large neighbourhood search-based heuristic which provides upper bounds for the proposed mathematical formulation, which proves effective in reducing solution times. We present a theoretical study of the problem, extending some known properties of the VRP with split deliveries, and an extensive computational analysis aimed at studying the strengths and weakness of the proposed strategy.

Constraint-Based Fleet Design Optimisation for Multi-compartment Split-Delivery Rich Vehicle Routing

We describe a large neighbourhood search (LNS) solver based on a constraint programming (CP) model for a real-world rich vehicle routing problem with compartments arising in the context of fuel delivery. Our solver supports both single-day and multi-day scenarios and a variety of real-world aspects including time window constraints, compatibility constraints, and split deliveries. It can be used both to plan the daily delivery operations, and to inform decisions on the long-term fleet composition. We show experimentally the viability of our approach.

A model-based genetic algorithm framework for constrained optimisation problems

Two major challenges are presented when applying genetic algorithms (GAs) to constrained optimisation problems: modelling and constraint handling. The field of constraint programming (CP) has enjoyed extensive research in both of these areas. CP frameworks have been devised which allow arbitrary problems to be readily modelled, and their constraints handled efficiently. Our work aims to combine the modelling and constraint handling of a state-of-the-art CP framework with the efficient population-based search of a GA. We present a new general hybrid CP / GA framework which can be used to solve any constrained optimisation problem that can be expressed using the language of constraints. The efficacy of this framework as a general heuristic for constrained optimisation problems is demonstrated through experimental results on a variety of classical combinatorial optimisation problems commonly found in the literature.

Planning Habitat Restoration with Genetic Algorithms

Conservation is an ethic of sustainable use of natural resources which focuses on the preservation of biodiversity. The term conservation planning encompasses the set of activities, typically carried out by conservation managers, that contribute to the attainment of this goal. Such activities can be preventive, such as the establishment of conservation reserves, or remedial, such as the displacement (or offsetting) of the species to be protected or the culling of invasive species. This last technique is often referred to as habitat restoration and, because of its lower impact on economic activities, is becoming more and more popular among conservation managers. In this paper we present the original formulation of the habitat restoration planning (HRP) problem, which captures some of the decisions and constraints faced by conservation managers in the context of habitat restoration. Example scenarios are drawn from the insular Great Barrier Reef (QLD) and Pilbara (WA) regions of Australia. In addition to the problem formulation, we describe an optimisation solver for the HRP, based on genetic algorithms (GAs), we discuss the preliminary results obtained by our solver, and we outline the current and future directions for the project.