Hyper-heuristics is a new field of optimisation which has recently emerged and is receiving growing exposure in the research community and literature. Hyper-heuristics are optimisation methods which are designed with a high level of abstraction from any one specific problem or class of problems and therefore are more generally applicable than specialised meta-heuristic and heuristic methods. Instead of being designed to solve a specific real-world problem, hyper-heuristics are designed to solve the problem of heuristic generation and selection. As such, hyper-heuristics can be thought of as methods for optimising the operations of an optimisation process which finds good solutions to a problem as a by-product. This approach has been shown to be very effective and in some cases provides improvement in search performance as well as reducing the burden associated with tailoring meta-heuristics which is often required when solving new problems. In this thesis, the hypothesis that hyper-heuristics can be competitively applied to real-world multi-objective optimisation problems such as the water distribution design problem is tested. Although many single-objective hyper-heuristics have been proposed in the literature, only a few multi-objective methods have been proposed. This thesis explores two different novel multi-objective hyper-heuristics: one designed for generating new specialised heuristics; and one designed for solving the online selection of heuristics. Firstly, the behaviour of a set of heuristics is explored to create a base understanding of different heuristic behavioural traits in order to better understand the hyper-heuristic behaviours and dynamics later in the study. Both approaches are tested on a range of benchmark optimisation problems and finally applied to real-world instances of the water distribution network design problem where the selective hyper-heuristics is demonstrated as being very effective at solving this difficult problem. Furthermore, the thesis demonstrates how heuristic selection can be improved by incorporating a greater level of information about heuristic performance, namely the historical joint performance of different heuristics, and shows that exploiting this sequencing information in heuristic selection can produce highly competitive results.

Decision-tree induction is one of the most employed methods to extract knowledge from data. There are several distinct strategies for inducing decision trees from data, each one presenting advantages and disadvantages according to its corresponding inductive bias. These strategies have been continuously improved by researchers over the last 40 years. This thesis, following recent breakthroughs in the automatic design of machine learning algorithms, proposes to automatically generate decision-tree induction algorithms. Our proposed approach, namely HEAD-DT, is based on the evolutionary algorithms paradigm, which improves solutions based on metaphors of biological processes. HEAD-DT works over several manually-designed decision-tree components and combines the most suitable components for the task at hand. It can operate according to two different frameworks: i) evolving algorithms tailored to one single data set (specific framework); and ii) evolving algorithms from multiple data sets (general framework). The specific framework aims at generating one decision-tree algorithm per data set, so the resulting algorithm does not need to generalise beyond its target data set. The general framework has a more ambitious goal, which is to generate a single decision-tree algorithm capable of being effectively applied to several data sets. The specific framework is tested over 20 UCI data sets, and results show that HEAD-DTs specific algorithms outperform algorithms like CART and C4.5 with statistical significance. The general framework, in turn, is executed under two different scenarios: i) designing a domain-specific algorithm; and ii) designing a robust domain-free algorithm. The first scenario is tested over 35 microarray gene expression data sets, and results show that HEAD-DTs algorithms consistently outperform C4.5 and CART in different experimental configurations. The second scenario is tested over 67 UCI data sets, and HEAD-DTs algorithms were shown to be competitive with C4.5 and CART. Nevertheless, we show that HEAD-DT is prone to a special case of overfitting when it is executed under the second scenario of the general framework, and we point to possible alternatives for solving this problem. Finally, we perform an extensive experiment for evaluating the best single-objective fitness function for HEAD-DT, combining 5 classification performance measures with three aggregation schemes. We evaluate the 15 fitness functions in 67 UCI data sets, and the best of them are employed to generate algorithms tailored to balanced and imbalanced data. Results show that the automatically-designed algorithms outperform CART and C4.5 with statistical significance, indicating that HEAD-DT is also capable of generating custom algorithms for data with a particular kind of statistical profile / Árvores de decisão são amplamente utilizadas como estratégia para extração de conhecimento de dados. Existem muitas estratégias diferentes para indução de árvores de decisão, cada qual com suas vantagens e desvantagens tendo em vista seu bias indutivo. Tais estratégias têm sido continuamente melhoradas por pesquisadores nos últimos 40 anos. Esta tese, em sintonia com recentes descobertas no campo de projeto automático de algoritmos de aprendizado de máquina, propõe a geração automática de algoritmos de indução de árvores de decisão. A abordagem proposta, chamada de HEAD-DT, é baseada no paradigma de algoritmos evolutivos. HEAD-DT evolui componentes de árvores de decisão que foram manualmente codificados e os combina da forma mais adequada ao problema em questão. HEAD-DT funciona conforme dois diferentes frameworks: i) evolução de algoritmos customizados para uma única base de dados (framework específico); e ii) evolução de algoritmos a partir de múltiplas bases (framework geral). O framework específico tem por objetivo gerar um algoritmo por base de dados, de forma que o algoritmo projetado não necessite de poder de generalização que vá além da base alvo. O framework geral tem um objetivo mais ambicioso: gerar um único algoritmo capaz de ser efetivamente executado em várias bases de dados. O framework específico é testado em 20 bases públicas da UCI, e os resultados mostram que os algoritmos específicos gerados por HEAD-DT apresentam desempenho preditivo significativamente melhor do que algoritmos como CART e C4.5. O framework geral é executado em dois cenários diferentes: i) projeto de algoritmo específico a um domínio de aplicação; e ii) projeto de um algoritmo livre-de-domínio, robusto a bases distintas. O primeiro cenário é testado em 35 bases de expressão gênica, e os resultados mostram que o algoritmo gerado por HEAD-DT consistentemente supera CART e C4.5 em diferentes configurações experimentais. O segundo cenário é testado em 67 bases de dados da UCI, e os resultados mostram que o algoritmo gerado por HEAD-DT é competitivo com CART e C4.5. No entanto, é mostrado que HEAD-DT é vulnerável a um caso particular de overfitting quando executado sobre o segundo cenário do framework geral, e indica-se assim possíveis soluções para tal problema. Por fim, é realizado uma análise detalhada para avaliação de diferentes funções de fitness de HEAD-DT, onde 5 medidas de desempenho são combinadas com três esquemas de agregação. As 15 versões são avaliadas em 67 bases da UCI e as melhores versões são utilizadas para geração de algoritmos customizados para bases balanceadas e desbalanceadas. Os resultados mostram que os algoritmos gerados por HEAD-DT apresentam desempenho preditivo significativamente melhor que CART e C4.5, em uma clara indicação que HEAD-DT também é capaz de gerar algoritmos customizados para certo perfil estatístico dos dados de classificação

Decision-tree induction is one of the most employed methods to extract knowledge from data. There are several distinct strategies for inducing decision trees from data, each one presenting advantages and disadvantages according to its corresponding inductive bias. These strategies have been continuously improved by researchers over the last 40 years. This thesis, following recent breakthroughs in the automatic design of machine learning algorithms, proposes to automatically generate decision-tree induction algorithms. Our proposed approach, namely HEAD-DT, is based on the evolutionary algorithms paradigm, which improves solutions based on metaphors of biological processes. HEAD-DT works over several manually-designed decision-tree components and combines the most suitable components for the task at hand. It can operate according to two different frameworks: i) evolving algorithms tailored to one single data set (specific framework); and ii) evolving algorithms from multiple data sets (general framework). The specific framework aims at generating one decision-tree algorithm per data set, so the resulting algorithm does not need to generalise beyond its target data set. The general framework has a more ambitious goal, which is to generate a single decision-tree algorithm capable of being effectively applied to several data sets. The specific framework is tested over 20 UCI data sets, and results show that HEAD-DTs specific algorithms outperform algorithms like CART and C4.5 with statistical significance. The general framework, in turn, is executed under two different scenarios: i) designing a domain-specific algorithm; and ii) designing a robust domain-free algorithm. The first scenario is tested over 35 microarray gene expression data sets, and results show that HEAD-DTs algorithms consistently outperform C4.5 and CART in different experimental configurations. The second scenario is tested over 67 UCI data sets, and HEAD-DTs algorithms were shown to be competitive with C4.5 and CART. Nevertheless, we show that HEAD-DT is prone to a special case of overfitting when it is executed under the second scenario of the general framework, and we point to possible alternatives for solving this problem. Finally, we perform an extensive experiment for evaluating the best single-objective fitness function for HEAD-DT, combining 5 classification performance measures with three aggregation schemes. We evaluate the 15 fitness functions in 67 UCI data sets, and the best of them are employed to generate algorithms tailored to balanced and imbalanced data. Results show that the automatically-designed algorithms outperform CART and C4.5 with statistical significance, indicating that HEAD-DT is also capable of generating custom algorithms for data with a particular kind of statistical profile / Árvores de decisão são amplamente utilizadas como estratégia para extração de conhecimento de dados. Existem muitas estratégias diferentes para indução de árvores de decisão, cada qual com suas vantagens e desvantagens tendo em vista seu bias indutivo. Tais estratégias têm sido continuamente melhoradas por pesquisadores nos últimos 40 anos. Esta tese, em sintonia com recentes descobertas no campo de projeto automático de algoritmos de aprendizado de máquina, propõe a geração automática de algoritmos de indução de árvores de decisão. A abordagem proposta, chamada de HEAD-DT, é baseada no paradigma de algoritmos evolutivos. HEAD-DT evolui componentes de árvores de decisão que foram manualmente codificados e os combina da forma mais adequada ao problema em questão. HEAD-DT funciona conforme dois diferentes frameworks: i) evolução de algoritmos customizados para uma única base de dados (framework específico); e ii) evolução de algoritmos a partir de múltiplas bases (framework geral). O framework específico tem por objetivo gerar um algoritmo por base de dados, de forma que o algoritmo projetado não necessite de poder de generalização que vá além da base alvo. O framework geral tem um objetivo mais ambicioso: gerar um único algoritmo capaz de ser efetivamente executado em várias bases de dados. O framework específico é testado em 20 bases públicas da UCI, e os resultados mostram que os algoritmos específicos gerados por HEAD-DT apresentam desempenho preditivo significativamente melhor do que algoritmos como CART e C4.5. O framework geral é executado em dois cenários diferentes: i) projeto de algoritmo específico a um domínio de aplicação; e ii) projeto de um algoritmo livre-de-domínio, robusto a bases distintas. O primeiro cenário é testado em 35 bases de expressão gênica, e os resultados mostram que o algoritmo gerado por HEAD-DT consistentemente supera CART e C4.5 em diferentes configurações experimentais. O segundo cenário é testado em 67 bases de dados da UCI, e os resultados mostram que o algoritmo gerado por HEAD-DT é competitivo com CART e C4.5. No entanto, é mostrado que HEAD-DT é vulnerável a um caso particular de overfitting quando executado sobre o segundo cenário do framework geral, e indica-se assim possíveis soluções para tal problema. Por fim, é realizado uma análise detalhada para avaliação de diferentes funções de fitness de HEAD-DT, onde 5 medidas de desempenho são combinadas com três esquemas de agregação. As 15 versões são avaliadas em 67 bases da UCI e as melhores versões são utilizadas para geração de algoritmos customizados para bases balanceadas e desbalanceadas. Os resultados mostram que os algoritmos gerados por HEAD-DT apresentam desempenho preditivo significativamente melhor que CART e C4.5, em uma clara indicação que HEAD-DT também é capaz de gerar algoritmos customizados para certo perfil estatístico dos dados de classificação

Supply chain optimization is one of the key components for the effective management of a company with a complex manufacturing process and distribution network. Companies with a global presence in particular are motivated to optimize their distribution plans in order to keep their operating costs low and competitive. Changing condition in the global market and volatile energy prices increase the need for an automatic decision and optimization tool. In recent years, many techniques and applications have been proposed to address the problem of supply chain optimization. However, such techniques are often too problemspecific or too knowledge-intensive to be implemented as in-expensive, and easy-to-use computer system. The effort required to implement an optimization system for a new instance of the problem appears to be quite significant. The development process necessitates the involvement of expert personnel and the level of automation is low. The aim of this project is to develop a set of strategies capable of increasing the level of automation when developing a new optimization system. An increased level of automation is achieved by focusing on three areas: multi-objective optimization, optimization algorithm usability, and optimization model design. A literature review highlighted the great level of interest for the problem of multiobjective optimization in the research community. However, the review emphasized a lack of standardization in the area and insufficient understanding of the relationship between multi-objective strategies and problems. Experts in the area of optimization and artificial intelligence are interested in improving the usability of the most recent optimization algorithms. They stated the concern that the large number of variants and parameters, which characterizes such algorithms, affect their potential applicability in real-world environments. Such characteristics are seen as the root cause for the low success of the most recent optimization algorithms in industrial applications. Crucial task for the development of an optimization system is the design of the optimization model. Such task is one of the most complex in the development process, however, it is still performed mostly manually. The importance and the complexity of the task strongly suggest the development of tools to aid the design of optimization models. In order to address such challenges, first the problem of multi-objective optimization is considered and the most widely adopted techniques to solve it are identified. Such techniques are analyzed and described in details to increase the level of standardization in the area. Empirical evidences are highlighted to suggest what type of relationship exists between strategies and problem instances. Regarding the optimization algorithm, a classification method is proposed to improve its usability and computational requirement by automatically tuning one of its key parameters, the termination condition. The algorithm understands the problem complexity and automatically assigns the best termination condition to minimize runtime. The runtime of the optimization system has been reduced by more than 60%. Arguably, the usability of the algorithm has been improved as well, as one of the key configuration tasks can now be completed automatically. Finally, a system is presented to aid the definition of the optimization model through regression analysis. The purpose of the method is to gather as much knowledge about the problem as possible so that the task of the optimization model definition requires a lower user involvement. The application of the proposed algorithm is estimated that could have saved almost 1000 man-weeks to complete the project. The developed strategies have been applied to the problem of Caterpillar’s global supply chain optimization. This thesis describes also the process of developing an optimization system for Caterpillar and highlights the challenges and research opportunities identified while undertaking this work. This thesis describes the optimization model designed for Caterpillar’s supply chain and the implementation details of the Ant Colony System, the algorithm selected to optimize the supply chain. The system is now used to design the distribution plans of more than 7,000 products. The system improved Caterpillar’s marginal profit on such products by a factor of 4.6% on average.

The Vehicle Routing Problem has been around for more than 50 years and has been of major interest to the operations research community. The VRP pose a complex problem with major benefits for the industry. In every supply chain transportation occurs between customers and suppliers. In this thesis, we analyze the use of a multiple pheromone trial in using Ant Systems to solve the VRP. The goal is to find a reasonable solution for data environments of derivatives of the basic VRP. An adaptive object model approach is followed to allow for additional constraints and customizable cost functions. A parallel method is used to improve speed and traversing the solution space. The Ant System is applied to the local search operations as well as the data objects. The Tabu Search method is used in the local search part of the solution. The study succeeds in allowing for all of the key performance indicators, i.e. efficiency, effectiveness, alignment, agility and integration for an IT system, where the traditional research on a VRP algorithm only focuses on the first two. / Thesis (PhD)--University of Pretoria, 2010. / Industrial and Systems Engineering / unrestricted

The importance of balance between exploration and exploitation plays a crucial role while solving combinatorial optimization problems. This balance is reached by two general techniques: by using an appropriate problem solver and by setting its proper parameters. Both problems were widely studied in the past and the research process continues up until now. The latest studies in the field of automated machine learning propose merging both problems, solving them at design time, and later strengthening the results at runtime. To the best of our knowledge, the generalized approach for solving the parameter setting problem in heuristic solvers has not yet been proposed. Therefore, the concept of merging heuristic selection and parameter control have not been introduced. In this thesis, we propose an approach for generic parameter control in meta-heuristics by means of reinforcement learning (RL). Making a step further, we suggest a technique for merging the heuristic selection and parameter control problems and solving them at runtime using RL-based hyper-heuristic. The evaluation of the proposed parameter control technique on a symmetric traveling salesman problem (TSP) revealed its applicability by reaching the performance of tuned in online and used in isolation underlying meta-heuristic. Our approach provides the results on par with the best underlying heuristics with tuned parameters.:1 Introduction 1 1.1 Motivation 1 1.2 Research objective 2 1.3 Solution overview 2 2 Background and RelatedWork Analysis 3 2.1 Optimization Problems and their Solvers 3 2.2 Heuristic Solvers for Optimization Problems 9 2.3 Setting Algorithm Parameters 19 2.4 Combined Algorithm Selection and Hyper-Parameter Tuning Problem 27 2.5 Conclusion on Background and Related Work Analysis 28 3 Online Selection Hyper-Heuristic with Generic Parameter Control 31 3.1 Combined Parameter Control and Algorithm Selection Problem 31 3.2 Search Space Structure 32 3.3 Parameter Prediction Process 34 3.4 Low-Level Heuristics 35 3.5 Conclusion of Concept 36 4 Implementation Details 37 4.2 Search Space 40 4.3 Prediction Process 43 4.4 Low Level Heuristics 48 4.5 Conclusion 52 5 Evaluation 55 5.1 Optimization Problem 55 5.2 Environment Setup 56 5.3 Meta-heuristics Tuning 56 5.4 Concept Evaluation 60 5.5 Analysis of HH-PC Settings 74 5.6 Conclusion 79 6 Conclusion 81 7 FutureWork 83 7.1 Prediction Process 83 7.2 Search Space 84 7.3 Evaluations and Benchmarks 84 Bibliography 87 A Evaluation Results 99 A.1 Results in Figures 99 A.2 Results in numbers 105