Discrete Optimization
Parallel local search algorithms for high school timetabling problems

https://doi.org/10.1016/j.ejor.2017.07.029Get rights and content

Highlights

  • We design parallel metaheuristics for high school timetabling problems.

  • Several different metaheuristics are used as agents.

  • Features such as cooperation, intensification and diversification are analyzed.

  • Strong elitism and exchange of information among threads improve the search.

  • Our best algorithm outperforms existing state-of-the-art algorithms.

Abstract

High school timetabling consists in assigning meetings between classes and teachers, with the goal of minimizing the violation of specific soft requirements. This family of problems has been frequently considered in the literature, but few strategies employing parallelism have been proposed. In this exploratory study, we consider two different parallel frameworks and present a thorough computational study in order to understand algorithmic decisions that are closely related to performance. Our best algorithm outperforms state-of-the-art algorithms for variants of the problem considered, indicating both the efficiency and the flexibility of the method.

Introduction

Educational timetabling problems consist in scheduling encounters between teachers (or exams) and students. Specific situations give origin to different problems, with a richness of characteristics. The scientific literature has branched this family of problems in three main sub-areas: University course timetabling Di Gaspero, McCollum, and Schaerf (2007); Lewis (2008); Lewis, Paechter, and McCollum (2007), examination timetabling (McCollum, McMullan, Burke, Parkes, Qu, 2007, Qu, Burke, McCollum, Merlot, Lee, 2009) and high school timetabling (De Werra, 1985, Pillay, 2014, Post, Ahmadi, Daskalaki, Kingston, Kyngas, Nurmi, Ranson, 2012, Schaerf, 1999). Each of these three families contains their own set of specific constraints.

In this article, we address the high school timetabling problem (HSTP). A restricted decision version of the HSTP was shown to be NP-complete by polynomial reduction of the 3-SAT problem (Even, Itai, & Shamir, 1975). Instances originating from practical contexts naturally extend the requirements of this idealized problem. Extra requirements are usually associated with pedagogical preferences and social/cultural particularities. Examples of these extra requirements are the need to assign double lessons (two consecutive lessons) for some classes subjects and the need to obtain compact schedules for teachers (Pillay, 2014).

The HSTP is usually modeled by mixed-integer programming (MIP) models (Al-Yakoob, Sherali, 2015, Dorneles, de Araújo, Buriol, 2014, Dorneles, de Araújo, Buriol, 2017, Kristiansen, Sørensen, Stidsen, 2015, Santos, Uchoa, Ochi, Maculan, 2012, Sørensen, Dahms, 2014). The resolution of these models for most medium and large practical instances are known to be still a challenge for actual black-box MIP solvers, which may be why most of the literature is concerned with metaheuristic techniques (Pillay, 2014). The main goal in these metaheuristic studies is to develop strategies to generate good quality solutions within reasonable computational efforts. Recently proposed metaheuristics can be found in Dorneles et al. (2014); Fonseca and Santos (2014); Fonseca, Santos, and Carrano (2016a); Fonseca, Santos, Toffolo, Brito, and Souza (2016c); Saviniec, Constantino, Romão, and Santos (2013). These heuristics are shown to be efficient in the sense that near-optimal solutions are consistently found for different input instances.

In this article, we aim to investigate parallel metaheuristics for the HSTP, in which different solution methods (agents) are run concurrently in different processor threads. The motivation is two-fold. On one hand, the availability of multi-processor machines and appropriate coding schemes make the use of parallel algorithms more accessible than ever. On the other hand, although the literature is abundant in parallel solution methods for similar complex problems (Subramanian, Drummond, Bentes, Ochi, Farias, 2010, Bożejko, Pempera, Smutnicki, 2013, Sánchez-Oro, Sevaux, Rossi, Martí, Duarte, 2015, Luque, Alba, 2015, e.g.), very little attention has been given to the design of parallel methods for the HSTP (Abramson, 1991, Abramson, Abela, 1992, Srndic, Pandzo, Dervisevic, Konjicija, 2009). The reader is referred to Section 3 for a closer look at these contributions.

As an exploratory study, our goal is to investigate a number of questions related to the design of parallel metaheuristics in the context of the HSTP. The main research questions can be summarized as: (1) What are good parallel strategies for the HSTP? (2) Can these strategies improve the performance of stand-alone metaheuristics? (3) Are the proposed parallel metaheuristics competitive with state-of-the-art methods?

In order to provide some insight into these questions, we propose a thorough computational study that tests all parallelization schemes resulting from combinations of the following characteristics:

Agents cooperation: we wish to test the effect of allowing the agents to cooperate, by sharing good solutions among threads.

Diversification: we wish to test the effect of using all agents as search intensification mechanisms or allowing at least one agent to diversify the search.

Agents diversity: we wish to test homogeneous and heterogeneous algorithms. An algorithm is homogeneous when all threads execute the same metaheuristic, and heterogeneous otherwise.

As stand-alone metaheuristics, we use the iterated local search from Saviniec et al. (2013) and adapt versions of tabu search, simulated annealing, and late acceptance strategy. We compare the performance of the different parallel algorithm variants against each other, against the stand-alone metaheuristics and against two state-of-the-art algorithms for close variants of the problem Dorneles et al. (2014); Fonseca et al. (2016a).

The remainder of this paper is organized as follows. Section 2 describes the problem both in plain language and with the help of a formal mixed-integer programming formulation. Section 3 gives an introduction to different parallelization schemes and reviews the existing parallel algorithms for the HSTP. Section 4 explains our parallel metaheuristics. Section 5 presents our computational experiments and Section 6 concludes this paper with final remarks and suggestions for further investigations. An appendix completes this article with details on the used metaheuristics and their parameter setting procedures.

Section snippets

The high school timetabling problem

We focus on a real HSTP motivated by Brazilian high schools timetabling rules. In this context, classes are disjoint groups of students enrolled in the same set of subjects (mathematics, chemistry, e.g.) and with no free periods during the week. The school also has a set of teachers in its workforce and the goal of the problem is to obtain weekly timetable specifying the schedule of meetings for class/teacher pairs.

Definition 1 HSTP instance

A HSTP instance is the input data for the timetabling construction in a

Existing parallel algorithms for the HSTP

Metaheuristics can be classified into population-based (PB) and trajectory-based (TB) methods (Alba, Luque, & Nesmachnow, 2013). Population-based metaheuristics are characterized by keeping a pool of solutions (genetic algorithms, ant colony optimization, and particle swarm optimization, e.g.). The method starts with an initial population and employs, at each step, stochastic operators to evolve toward better quality populations. According to Alba et al. (2013), two classical parallel algorithm

Proposed parallel metaheuristics

The proposed algorithms follow the trajectory-based parallel multi-start (TB-PMS) scheme described in the previous section and employ manager/workers strategies with shared memories. We propose the use of two main strategies. The first is based on central memory (Crainic, Gendreau, Hansen, & Mladenović, 2004) and the second is based on diversification and intensification memories (Jin, Crainic, & Løkketangen, 2014).

Computational experiments

In this section, we conduct a thorough computational study in order to evaluate the performance of the proposed parallel methods. All algorithms were coded in C++ and compiled with GNU Compiler Collection 4.4.7. The experiments were made on a server running Red Hat Enterprise Linux 6.5. The hardware is composed of two CPU Intel Xeon E5-2680v2 (2.8 gigahertz) and 128 gigabytes of RAM. To implement parallelism we employed the POSIX Threads Library (Pthreads) available in the GNU Compiler

Conclusions

In this paper, we conducted an extensive study with parallel trajectory-based metaheuristics for high school timetabling problems. Our study analyzed several aspects related to the design of these algorithms. In particular, we studied the homogeneity/heterogeneity of agents, the influence of intensification and diversification memories, and the frequency of information exchange among agents.

Our most efficient algorithm was obtained with the inclusion of a diversification memory, a very elitist

Acknowledgments

This research was supported by FAPESP-Brazil. Grants: 2013/13563-3 and 2015/10032-2. The authors would like to thank Professor George H. G. Fonseca for sharing the code of GOAL and also, the anonymous reviewers for their useful comments in our paper.

References (45)

  • L. Saviniec et al.

    Effective local search algorithms for high school timetabling problems

    Applied Soft Computing

    (2017)
  • M. Sørensen et al.

    A two-stage decomposition of high school timetabling applied to cases in denmark

    Computers & Operations Research

    (2014)
  • A. Subramanian et al.

    A parallel heuristic for the vehicle routing problem with simultaneous pickup and delivery

    Computers & Operations Research

    (2010)
  • D. Abramson

    Constructing school timetables using simulated annealing: sequential and parallel algorithms

    Management science

    (1991)
  • D. Abramson et al.

    A parallel genetic algorithm for solving the school timetabling problem

    Proceedings of the 15 Australian computer science conference

    (1992)
  • E. Alba et al.

    Parallel metaheuristics: Recent advances and new trends

    International Transactions in Operational Research

    (2013)
  • R. Burget et al.

    Teacher-oriented fairness in course timetabling

    Proceedings of the 11th international conference on the practice and theory of automated timetabling (patat 2016)

    (2016)
  • E.K. Burke et al.

    A late acceptance strategy in hill-climbing for exam timetabling problems

    Proceedings of the 7th conference on the practice and theory of automated timetabling, montreal, canada

    (2008)
  • T.G. Crainic et al.

    Cooperative parallel variable neighborhood search for the p-median

    Journal of Heuristics

    (2004)
  • L. Di Gaspero et al.

    The second international timetabling competition (ITC-2007): Curriculum-based course timetabling (track 3)

    Technical Report

    (2007)
  • S. Even et al.

    On the complexity of timetable and multi-commodity flow problems

    16th annual symposium on foundations of computer science

    (1975)
  • G.H. Fonseca et al.

    Late acceptance hill-climbing for high school timetabling

    Journal of Scheduling

    (2016)
  • Cited by (25)

    • High quality timetables for Italian schools

      2022, Computers and Operations Research
    • A survey of the state-of-the-art of optimisation methodologies in school timetabling problems

      2021, Expert Systems with Applications
      Citation Excerpt :

      The highest degree first heuristic was used for the ordering of the coloured edges, and the computational results demonstrated that it was effective. Saviniec et al. (2018) proposed a parallel local search in solving Brazilian high school timetabling problems. The algorithm employed trajectory-based parallel multi-start (ran with several asynchronous threads of trajectory-based methods simultaneously) and manager/workers strategies with shared memories.

    • Handling the pseudo pilot assignment problem in air traffic control training by using NASA TLX

      2020, Journal of Air Transport Management
      Citation Excerpt :

      In their study, the preferences of the lecturers were also taken into consideration. For more studies about teacher-course assignment, see Badri (1996), Ozdemir and Gasimov (2004), for teacher-class assignment: (Hultberg and Cardoso (1997), Asratian and De Werra (2002), Carrasco and Pato (2004), for student-classroom assignment: Vermuyten et al. (2016), and for general timetabling De Causmaecker et al. (2009), Saviniec et al. (2018), Sunday et al. (2018), Demirović and Stuckey (2018), Saviniec et al. (2020) and Tassopoulos et al. (2020). The department of ATC in the faculty of aeronautics and astronautics at ESTU has six types of simulation courses.

    View all citing articles on Scopus
    View full text