Parameter tuning for cross-domain search

Gumus, Duriye Betul (2020) Parameter tuning for cross-domain search. PhD thesis, University of Nottingham.

[img] PDF (Thesis - as examined) - Repository staff only - Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader
Download (1MB)


Metaheuristics usually have algorithmic parameters whose initial settings can influence their search behaviour and arbitrarily setting these values often leads to poor performance. Parameter tuning, i.e. determining the best initial parameter values, is a challenging and time-consuming task, but is one which is crucial to obtaining improved meta-heuristic performance. Memetic algorithms, which hybridise evolutionary algorithms with local search, are well-known metaheuristics for solving combinatorial optimisation problems. There are various methods available for parameter tuning and these have been widely used to tune the parameters of metaheuristics for individual optimisation problems.

There is a growing interest towards designing general-purpose problem-independent search methodologies which produce good quality solutions for cross-domain search; which means that they consider a broad range of optimisation problems rather than being specialised for a solution method for a single problem. Metaheuristics can be used as a high-level cross-domain search method, utilising the available low level heuristics for each of the problem domains to modify a solution. The cross-domain search methods have ‘algorithm parameters’ to control the behaviour of the algorithms, while the low level heuristics have ‘heuristic parameters’ to determine the extent of the change that they make when used. Previous cross-domain search methods in the literature generally either fixed these parameters to some values based on previous experience or adapted them to the specific instances being solved. However, there has been a lack of extensive research in tuning these parameters for cross-domain search.

In this thesis, two parameter tuning methods are adapted as cross-domain parameter tuning approaches. The parameters of a steady state memetic algorithm (SSMA) and the low level heuristics which are used in this algorithm are tuned across nine problem domains using the training instances from only a limited number of problems. One of the issues with the cross-domain parameter tuning is that the instances are characteristically different in nature and that their objective function values are often not comparable; therefore, a Formula 1 ranking based Taguchi method is proposed as the first approach. Four strategies are also presented to illustrate and contrast the capability of an F-Race parameter tuning method for cross-domain search. Due to the potential for excessive runtime for these problems, an analysis of parameter tuning with reduced computational time budget is also presented. The empirical results show that both of the methods managed to find good parameter settings which outperform the untuned SSMA as well as many of the previous CHeSC2011 competitors. Moreover, comparisons of the two methods show that Taguchi method found the same best configuration more easily and consistently with different run time budgets, while F-Race identified different configurations as the best one in some of the strategies, depending on the configuration budget and in what order instances are introduced to the tuning process.

Item Type: Thesis (University of Nottingham only) (PhD)
Supervisors: Ozcan, Ender
Atkin, Jason
Keywords: metaheuristics, search, algorithms
Subjects: Q Science > QA Mathematics > QA 75 Electronic computers. Computer science
Faculties/Schools: UK Campuses > Faculty of Science > School of Computer Science
Item ID: 59715
Depositing User: GUMUS, Duriye Betul
Date Deposited: 21 Jul 2020 04:40
Last Modified: 21 Jul 2020 04:40

Actions (Archive Staff Only)

Edit View Edit View