Hyper-heuristics

People use e-government applications, trade financial goods online, purchase online, book hotel rooms using mobile booking apps, and make decisions based on information given in organisational information systems. All of these decisions are impacted by the context in which they are made, as this statement demonstrates: What is chosen frequently relies on how the choice is presented. Designers of choice settings, sometimes known as choice architects, can take use of these heuristics and biases. In this paper various digital nudging research papers are studied. The goal of this study is to see if and how different digital nudges displayed on an e-commerce website impact the customer's psychological state. A variety of research reviews have been done in order to better understand the psychological heuristics impact based on digital nudging.

The apparel industry is a class of textile industry. Generally, the production scheduling problem in the apparel industry belongs to Flow Shop Scheduling Problems (FSSP). There are many algorithms/techniques/heuristics for solving FSSP. Two of them are the Palmer Algorithm and the Gupta Algorithm. Hyper-heuristic is a class of heuristics that enables to combine of some heuristics to produce a new heuristic. GPHH is a hyper-heuristic that is based on genetic programming that is proposed to solve FSSP [1]. This paper presents the development of a computer program that implements the GPHH. Some experiments have been conducted for measuring the performance of GPHH. From the experimental results, GPHH has shown a better performance than the Palmer Algorithm and Gupta Algorithm.

In an attempt to ensure good-quality printouts of our technical reports, from the supplied PDF files, we process to PDF using Acrobat Distiller. We encourage our authors to use outline fonts coupled with embedding of the used subset of all fonts (in either Truetype or Type 1 formats) except for the standard Acrobat typeface families of Times, Helvetica (Arial), Courier and Symbol. In the case of papers prepared using TEX or LATEX we endeavour to use subsetted Type 1 fonts, supplied by Y&Y Inc., for the Computer Modern, Lucida Bright and Mathtime families, rather than the public-domain Computer Modern bitmapped fonts. Note that the Y&Y font subsets are embedded under a site license issued by Y&Y Inc. For further details of site licensing and purchase of these fonts visit Abstract Hyper-heuristics comprise a set of approaches with the common goal of automating the design and tuning of heuristic methods to solve hard computational search problems. The main goal is to produce more generally applicable search methodologies. The term hyper-heuristic was coined in the early 2000's to refer to the idea of 'heuristics to choose heuristics'. However, the idea of automating the design of combined heuristics can be traced back to the early 1960's. With the incorporation of Genetic Programming into hyper-heuristic research, a new type of hyper-heuristics has emerged that we have termed 'heuristics to generate heuristics'. The distinguishing feature of hyper-heuristics is that they operate on a search space of heuristics (or heuristic components) rather than directly on the search space of solutions to the underlying problem, as is the case with most meta-heuristic approaches. This paper presents a literature survey of hyper-heuristics including their origin and intellectual roots, a detailed account of the main types of approaches, and an overview of some related areas. Current research trends and directions for future research are also discussed.

This thesis presents a programme of research which investigated a genetic programming hyper-heuristic methodology to automate the heuristic design process for one, two and three dimensional packing problems.

Traditionally, heuristic search methodologies operate on a space of potential solutions to a problem. In contrast, a hyper-heuristic is a heuristic which searches a space of heuristics, rather than a solution space directly. The majority of hyper-heuristic research papers, so far, have involved selecting a heuristic, or sequence of heuristics, from a set pre-defined by the practitioner. Less well studied are hyper-heuristics which can create new heuristics, from a set of potential components.

This thesis presents a genetic programming hyper-heuristic which makes it possible to automatically generate heuristics for a wide variety of packing problems. The genetic programming algorithm creates heuristics by intelligently combining components. The evolved heuristics are shown to be highly competitive with human created heuristics. The methodology is first applied to one dimensional bin packing, where the evolved heuristics are analysed to determine their quality, specialisation, robustness, and scalability. Importantly, it is shown that these heuristics are able to be reused on unseen problems. The methodology is then applied to the two dimensional packing problem to determine if automatic heuristic generation is possible for this domain. The three dimensional bin packing and knapsack problems are then addressed. It is shown that the genetic programming hyper-heuristic methodology can evolve human competitive heuristics, for the one, two, and three dimensional cases of both of these problems. No change of parameters or code is required between runs. This represents the first packing algorithm in the literature able to claim human competitive results in such a wide variety of packing domains.

We evolve heuristics to guide staged deepening search for the hard game of FreeCell, obtaining top-notch solvers for this NP-Complete, human-challenging puzzle. We rst devise several novel heuristic measures and then employ a Hillis-style coevolutionary genetic algorithm to nd ecient com-binations of these heuristics. Our results signicantly sur-pass the best published solver to date by three distinct mea-sures: 1) Number of search nodes is reduced by 87%; 2) time to solution is reduced by 93%; and 3) average solution length is reduced by 41%. Our top solver is the best published Free-Cell player to date, solving 98 % of the standard Microsoft 32K problem set, and also able to beat high-ranking human players.

Applying evolutionary algorithms to new problem domains
is an exercise in the art of parameter tuning and design
decisions. A great deal of work has investigated ways to
automate the tuning of various EA parameters such as population
size, mutation options, etc. However, genotypeto-phenotype
mappings have typically been considered too
complex to adapt automatically. We demonstrate a genetic
representation learning method that uses meta-evolution to
adapt a bitstring encoding for a synthetic class of real-valued
optimization problems. The genetic representation we learn
performs as well or better than a Gray code both on new
instances of the problem class it was trained on and on problem
types that it was not trained on

The literature shows that one, two and three dimensional bin packing and knapsack packing are difficult problems in Operational Research. Many techniques, including exact, heuristic, and metaheuristic approaches, have been investigated to solve these problems and it is often not clear which method to use when presented with a new instance. This paper presents an approach which is motivated by the goal of building computer systems which can design heuristic methods. The overall aim is to explore the possibilities for automating the heuristic design process.

We present a genetic programming system to automatically generate a good quality heuristic for each instance. It is not necessary to change the methodology depending on the problem type (one, two or three dimensional knapsack and bin packing problems), and it therefore has a level of generality unmatched by other systems in the literature. We carry out an extensive suite of experiments and compare with the best human designed heuristics in the literature. Note that our heuristic design methodology uses the same parameters for all the experiments.

The contribution of this paper is to present a more general packing methodology than those currently available, and to show that, by using this methodology, it is possible for a computer system to design heuristics which are competitive with the human designed heuristics from the literature. This represents the first packing algorithm in the literature able to claim human competitive results in such a wide variety of packing domains.

In this article, a multi-objective evolutionary framework to build selection hyper-heuristics for solving instances of the 2D bin packing problem is presented. The approach consists of a multi-objective evolutionary learning process, using specific tailored genetic operators, to produce sets of variable length rules representing hyper-heuristics. Each hyper-heuristic builds a solution to a given problem instance by sensing the state of the instance , and deciding which single heuristic to apply at each decision point. The hyper-heuristics consider the minimization of two conflicting objectives when building a solution: the number of bins used to accommodate the pieces and the total time required to do the job. The proposed framework integrates three well-studied multi-objective evolutionary algorithms to produce sets of Pareto-approximated hyper-heuristics: the Non-dominated Sorting Genetic Algorithm-II, the Strength Pareto Evolutionary Algorithm 2, and the Generalized Differential Evolution Algorithm 3. We conduct an extensive experimental analysis using a large set of 2D bin packing problem instances containing convex and non-convex irregular pieces, under many conditions, settings and using several performance metrics. The analysis assesses the ro-bustness and flexibility of the proposed approach, providing encouraging results when compared against a set of well-known baseline single heuristics.

Genetic programming approaches have been employed in the literature to automatically design constructive heuristics for cutting and packing problems. These heuristics obtain results superior to human created constructive heuristics, but they do not generally obtain results of the same quality as local search heuristics, which start from an initial solution and iteratively improve it. If local search heuristics can be successfully designed through evolution, in addition to a constructive heuristic which initialises the solution, then the quality of results which can be obtained by automatically generated algorithms can be significantly improved. This paper presents a grammatical evolution methodology which automatically designs good quality local search heuristics that maintain their performance on new problem instances.

The idea behind hyper-heuristics is to discover some combination of straightforward heuristics to solve a wide range of problems. To be worthwhile, such combination should outperform the single heuristics. This paper presents a GAbased method that produces general hyper-heuristics for the dynamic variable ordering within Constraint Satisfaction Problems. The GA uses a variable-length representation, which evolves combinations of condition-action rules producing hyper-heuristics after going through a learning process which includes training and testing phases. Such hyper-heuristics, when tested with a large set of benchmark problems, produce encouraging results for most of the cases. The testebed is composed of problems randomly generated using an algorithm proposed by Prosser .

Job Shop Scheduling problems have become popular because of their many industrial and practical applications. Among the many solving strategies for this problem, selection hyper-heuristics have attracted attention due to their promising results in this and similar optimization problems. A selection hyper-heuristic is a method that determines which heuristic to apply at given points of the problem throughout the solving process. Unfortunately, results from previous studies show that selection hyper-heuristics are not free from making wrong choices. Hence, this paper explores a novel way of improving selection hyper-heuristics by using neural networks that are trained with information from existing selection hyper-heuristics. These networks learn high-level patterns that result in improved performance concerning the hyper-heuristics they were generated from. At the end of the process, the neural networks work as hyper-heuristics that perform better than their original counterparts. The results presented in this paper confirm the idea that we can refine existing hyper-heuristics to the point of being able to defeat the best possible heuristic for each instance. For example, one of our experiments generated one hyper-heuristic that produced a schedule that reduced the makespan of the one obtained by a synthetic oracle by ten days.

One of the annual issues that has to be addressed in English football is producing a fixture schedule for the holiday periods that reduces the travel distance for the fans and players. This problem can be seen as a minimisation problem which must abide to the constraints set by the Football Association. In this study, the performance of selection hyper-heuristics is investigated as a solution methodology. Hyper-heuristics aim to automate the process of selecting and combining simpler heuristics to solve computational search problems. A selection hyper-heuristic stores a single candidate solution in memory and iteratively applies selected low level heuristics to improve it. The results show that the learning hyper-heuristics outperform some previously proposed approaches and solutions published by the Football Association.

A hyper-heuristic refers to a search method or a learning mechanism for selecting or generating heuristics to solve computational search problems. Operating at a level of abstraction above that of a metaheuristic, it can be seen as an algorithm that tries to find an appropriate solution method at a given decision point rather than a solution. This paper introduces a new hyper-heuristic that combines elements from reinforcement learning and tabu search. It is applied to solve two complex stochastic scheduling problems arising in mining, namely the stochastic open-pit mine production scheduling problem with one processing stream (SMPS) and one of its generalizations, SMPS with multiple processing streams and stockpiles (SMPS+). The performance of the new hyper-heuristic is assessed by comparing it to several solution methods from the literature: problem-specific algorithms tailored for the two problems addressed in the paper and general hyper-heuristics, which use only limited problem-specific information. The computational results indicate that not only is the proposed new hyper-heuristic approach superior to the other hyper-heuristics, but it also provides results that are comparable to or improve on the results obtained by the state-of-the-art problem-specific methods.

In my PhD. dissertation, I focus on action planning and
constrained discrete optimization. I try to introduce novel
approaches to the field of single-agent planning by com-
bining standard techniques with meta-heuristic optimiza-
tion, machine-learning algorithms, hyper-heuristics and al-
gorithm selection approaches.
Our main goal is to create new and flexible planning algo-
rithms which would be suited for a large variety of real-life
problems. Planning is a fundamental and difficult problem
in AI and any new results in this area are directly applica-
ble to many other fields. They can be used for single-agent
or multi-agent action selection in both competitive or coop-
erative environment and as we focus on optimization, our
techniques are suitable for real-life problems that arise in
robotics or transportation.

University course timetabling is a complex optimization
problem. There are many components like departments, faculties, rooms,
and students making the problem huge and difficult to solve. Each
component enforces a set of normally conflicting constraints on the
solution space. The problem will be solved if courses are scheduled in
different rooms and within some specific time slots such that a set of
constraints are satisfied. In this paper, a graph-based hyper-heuristic is
proposed to find a solution to the problem. This is a two tiers modeling
approach combining hyper heuristic with graph coloring technique. The
upper tier heuristic is used to select a suitable heuristic to find a feasible
solution on the lower tier. To find the suitability of the proposed
approach, it has been applied to a real world case. The proposed approach
was able to satisfy all the hard and soft constraints. Based on the research
findings, it can be concluded that a graph-based hyper heuristic approach
is a suitable and computationally efficient method to find a solution to
university course timetabling problem.

Hyper-heuristics are methodologies used to choose from a set of heuristics and decide which one to apply given some properties of the current instance. When solving a Constraint Satisfaction Problem, the order in which the variables are selected to be instantiated has implications in the complexity of the search. We propose a neural network hyper-heuristic approach for variable ordering within Constraint Satisfaction Problems. The first step in our approach requires to generate a pattern that maps any given instance, expressed in terms of constraint density and tightness, to one adequate heuristic. That pattern is later used to train various neural networks which represent hyper-heuristics. The results suggest that neural networks generated through this methodology represent a feasible alternative to code hyper-heuristic which exploit the strengths of the heuristics to minimise the cost of finding a solution.

We present a genetic programming (GP) system to evolve reusable heuristics for the 2-D strip packing problem. The evolved heuristics are constructive, and decide both which piece to pack next and where to place that piece, given the current partial solution. This paper contributes to a growing research area that represents a paradigm shift in search methodologies. Instead of using evolutionary computation to search a space of solutions, we employ it to search a space of heuristics for the problem. A key motivation is to investigate methods to automate the heuristic design process. It has been stated in the literature that humans are very good at identifying good building blocks for solution methods. However, the task of intelligently searching through all of the potential combinations of these components is better suited to a computer. With such tools at their disposal, heuristic designers are then free to commit more of their time to the creative process of determining good components, while the computer takes on some of the design process by intelligently combining these components. This paper shows that a GP hyper-heuristic can be employed to automatically generate human competitive heuristics in a very-well studied problem domain.

The current state of the art in hyper-heuristic research comprises a set of approaches that share the common goal of automating the design and adaptation of heuristic methods to solve hard computational search problems. The main goal is to produce more generally applicable search methodologies. In this chapter we present an overview of previous categorisations of hyper-heuristics and provide a unified classification and definition, which capture the work that is being undertaken in this field. We distinguish between two main hyper-heuristic categories: heuristic selection and heuristic generation. Some representative examples of each category are discussed in detail. Our goals are to clarify the main features of existing techniques and to suggest new directions for hyper-heuristic research.

Hyper-heuristics are a class of high-level search techniques which operate on a search space of heuristics rather than directly on a search space of solutions. Early hyperheuristics focussed on selecting and applying a low-level heuristic at each stage of a search. Recent trends in hyper-heuristic research have led to a number of approaches being developed to automatically generate new heuristics from a set of heuristic components. This work investigates the suitability of using genetic programming as a hyper-heuristic methodology to generate constructive heuristics to solve the multidimensional 0-1 knapsack problem. A population of heuristics to rank knapsack items are trained on a subset of test problems and then applied to unseen instances. The results over a set of standard benchmarks show that genetic programming can be used to generate constructive heuristics which yield human-competitive results.

Every year the English football authorities produce a set of fixtures for the four main divisions in England. Over the Christmas and New Year period every team has to play two fixtures; one being played at their home venue and the other at an opponent's venue. There are various other constraints that also have to be respected with the overall objective being to minimize the total distance travelled by all teams. In this paper, I formally define the problem, discuss the data collection that I have undertaken and present the algorithm (which is based on depth first search, followed by a local search) I have developed. Using data from four seasons, I show that I am able to produce better schedules than those currently used.

Hyper-heuristics are methodologies that allow us to selectively apply the most suitable heuristic given the properties of the problem at hand. They can be applied in CSP in different ways, but one way which has received attention in recent years is variable ordering by using hyper-heuristics. To select the next variable, a set of heuristics exist and the hyper-heuristic decides, considering the features that describe the instance at hand, which heuristic is more suitable to be applied at the moment. This paper explores a hyperheuristic model for variable ordering within CSP based on vector hyper-heuristics. Each hyper-heuristic is represented as a set of vectors that maps instance features to heuristics. These vector hyper-heuristics are constructed by going into a local search method that modifies the hyper-heuristics. The results suggest that the approach is able to combine the strengths of different heuristics to perform well on a wide range of instances and compensate for their weaknesses on specific instances, resulting in an improvement in the performance of the search compared against the heuristics applied in isolation.

The bin-packing problem is a well known NP-Hard optimisation problem, and, over the years, many heuristics have been developed to generate good quality solutions. This paper outlines a genetic programming system which evolves a heuristic that decides whether to put a piece in a bin when presented with the sum of the pieces already in the bin and the size of the piece that is about to be packed. This heuristic operates in a fixed framework that iterates through the open bins, applying the heuristic to each one, before deciding which bin to use. The best evolved programs emulate the functionality of the human designed ‘first-fit’ heuristic. Thus, the contribution of this paper is to demonstrate that genetic programming can be employed to automatically evolve bin packing heuristics which are the same as high quality heuristics which have been designed by humans.

Genetic programming approaches have previously been employed in the literature to evolve heuristics for various combinatorial optimisation problems. This paper presents a hyper-heuristic genetic programming methodology to evolve more sophisticated one dimensional bin packing heuristics than have been evolved previously. The heuristics have access to a memory, which allows them to make decisions with some knowledge of their potential future impact. In contrast to previously evolved heuristics for this problem, we show that these heuristics evolve to draw upon this memory in order to facilitate better planning, and improved packings. This fundamental difference enables an evolved heuristic to represent a dynamic packing strategy rather than a fixed packing strategy. A heuristic can change its behaviour depending on the characteristics of the pieces it has seen before, because it has evolved to draw upon its experience.

Maximizing the lifetime of wireless sensor networks (WSNs) is an important and challenging research problem. Properly scheduling the movements of mobile sinks to balance the energy consumption of wireless sensor network is one of the most effective approaches to prolong the lifetime of wireless sensor networks. However, the existing mobile sink scheduling methods either require a great amount of computational time or lack effectiveness in finding high-quality scheduling solutions. To address the above issues, this paper proposes a novel hyperheuristic framework, which can automatically construct high-level heuristics to schedule the sink movements and prolong the network lifetime. In the proposed framework, a set of low-level heuristics are defined as building blocks to construct high-level heuristics and a set of random networks with different features are designed for training. Further, a genetic programming algorithm is adopted to automatically evolve promising high-level heuristics based on the building blocks and the training networks. By using the genetic programming to evolve more effective heuristics and applying these heuristics in a greedy scheme, our proposed hyper-heuristic framework can prolong the network lifetime competitively with other methods, with small time consumption. A series of comprehensive experiments, including both static and dynamic networks, are designed. The simulation results have demonstrated that the proposed method can offer a very promising performance in terms of network lifetime and response time.

An important challenge within hyper-heuristic research is to design search methodologies that work well, not only across different instances of the same problem, but also across different problem domains. This article conducts an empirical study involving three different domains in combinatorial optimisation: bin packing, permutation flow shop and personnel scheduling. Using a common software interface (HyFlex), the same algorithms (high-level strategies or hyperheuristics) can be readily run on all of them. The study is intended as a proof of concept of the proposed interface and domain modules, as a benchmark for testing the generalisation abilities of heuristic search algorithms. Several algorithms and variants from the literature were implemented and tested. From them, the implementation of iterated local search produced the best overall performance. Interestingly, this is one of the most conceptually simple competing algorithms, its advantage as a robust algorithm is probably due to two factors: (i) the simple yet powerful exploration/exploitation balance achieved by systematically combining a perturbation followed by local search; and (ii) its parameter-less nature. We believe that the challenge is still open for the design of robust algorithms that can learn and adapt to the available low-level heuristics, and thus select and apply them accordingly.

This article presents a method based on the multi-objective evolutionary algorithm NSGA-II to approximate hyper-heuristics for solving irregular 2D cutting stock problems under multiple objectives. In this case, additionally to the traditional objective of minimizing the number of sheets used to fit a finite number of irregular pieces, the time required to perform the placement task is also minimized, leading to a bi-objective minimization problem with a tradeoff between the number of sheets and the time required for placing all pieces. We solve this problem using multi-objective hyper-heuristics (MOHHs), whose main idea consists of finding a set of simple heuristics which can be combined to find a general solution for a wide range of problems, where a single heuristic is applied depending on the current condition of the problem, instead of applying a unique single heuristic during the whole placement process. The MO-HHs are approximated after going through a learning process by mean of the NSGA-II, which evolves combinations of condition-action rules producing at the end a set of Pareto-optimal MOHHs. We tested the approximated MMOHHs on several sets of benchmark problems, having outstanding results for most of the cases.

Hyper-heuristics are a class of high-level search techniques which operate on a search space of heuristics rather than directly on a search space of solutions. Early hyper-heuristics focussed on selecting and applying a low-level heuristic at each stage of a search. Recent trends in hyper-heuristic
research have led to a number of approaches being developed to automatically generate new heuristics from a set of heuristic components. This work investigates the suitability of using genetic programming as a hyper-heuristic methodology to generate con-structive heuristics to solve the multidimensional 0-1 knapsack problem. A population of heuristics to rank knapsack items are trained on a subset of test problems and then applied to unseen instances. The results over a set of standard benchmarks show
that genetic programming can be used to generate constructive heuristics which yield human-competitive results.

This paper introduces a neuro-evolutionary approach to produce hyper-heuristics for the dynamic variable ordering for hard binary constraint satisfaction problems. The model uses a GA to evolve a population of neural networks architectures and parameters. For every cycle in the GA process, the new networks are trained using backpropagation. When the process is over, the best trained individual in the last population of neural networks represents the general hyperheuristic.

Hyper-heuristics are high level search methodologies that operate over a set of heuristics which operate directly on the problem domain. In one of the hyper-heuristic frameworks, the goal is automating the process of selecting a human-designed low level heuristic at each step to construct a solution for a given problem. Constraint Satisfaction Problems (CSP) are well know NP complete problems. In this study, behaviours of two variable ordering heuristics Max-Conflicts (MXC) and Saturation Degree (SD) with respect to various combinations of constraint density and tightness values are investigated in depth over a set of random CSP instances. The empirical results show that the performance of these two heuristics are somewhat complementary and they vary for changing constraint density and tightness value pairs. The outcome is used to design three hyper-heuristics using MXC and SD as low level heuristics to construct a solution for unseen CSP instances. It has been observed that these hyper-heuristics improve the performance of individual low level heuristics even further in terms of mean consistency checks for some CSP instances.

It is common to find that training of selection hyper-heuristics is done perturbatively. The process usually starts with a random selection module and iterates over a set of instances until finding appropriate values for such module. In this work, however, we present a model for creating selection hyper-heuristics constructively. To achieve so, we use a set of instances evolved for such a task. For each low-level heuristic, we evolved a set of problem instances that are more easily solvable by that particular heuristic (compared to the other ones). Each group contains instances easily solvable with the corresponding heuristic but not with the remaining ones. Thus, our model creates its selector by calculating the centroid of each group. For doing so, the model defines a rule that maps said centroid to one corresponding action (in this case, a low-level heuristic). To test our approach, we select the one-dimensional Bin Packing Problem and set four target performance levels (which we refer to as deltas) for instance generation. Then, we analyze all the possible combinations of deltas. We study how performance of the generated hyper-heuristics shift when they are created using a different number of instances. Our data shows the feasibility of creating a hyper-heuristic under the stated conditions. Effectiveness of the model depends on the deltas used though we observed that higher deltas are useful while lower deltas are not. For example, when considering a delta level of 2.0, our method produced hyper-heuristics with an accumulated average waste 12% lower than that of the best heuristic. But, for a delta level of 0.5, it became impossible to outperform the heuristics.

An important challenge within hyper-heuristic research is to design search methodologies that work well, not only across different instances of the same problem, but also across different problem domains. This article conducts an empirical study involving three different domains in combinatorial optimisation: bin packing, permutation flow shop and personnel scheduling. Using a common software interface (HyFlex), the same algorithms (high-level strategies or hyperheuristics) can be readily run on all of them. The study is intended as a proof of concept of the proposed interface and domain modules, as a benchmark for testing the generalisation abilities of heuristic search algorithms. Several algorithms and variants from the literature were implemented and tested. From them, the implementation of iterated local search produced the best overall performance. Interestingly, this is one of the most conceptually simple competing algorithms, its advantage as a robust algorithm is probably due to two factors: (i) the simple yet powerful exploration/exploitation balance achieved by systematically combining a perturbation followed by local search; and (ii) its parameter-less nature. We believe that the challenge is still open for the design of robust algorithms that can learn and adapt to the available low-level heuristics, and thus select and apply them accordingly.

Hyper-heuristics comprise a set of approaches that are motivated (at least in part) by the goal of automating the design of heuristic methods to solve hard computational search problems. An underlying strategic research challenge is to develop more generally applicable search methodologies. The term hyper-heuristic is relatively new; it was first used in 2000 to describe heuristics to choose heuristics in the context of combinatorial optimisation. However, the idea of automating the design of heuristics is not new; it can be traced back to the 1960s. The definition of hyper-heuristics has been recently extended to refer to a search method or learning mechanism for selecting or generating heuristics to solve computational search problems. Two main hyper-heuristic categories can be considered: heuristic selection and heuristic generation. The distinguishing feature of hyper-heuristics is that they operate on a search space of heuristics (or heuristic components) rather than directly on the search space of solutions to the underlying problem that is being addressed. This paper presents a critical discussion of the scientific literature on hyper-heuristics including their origin and intellectual roots, a detailed account of the main types of approaches, and an overview of some related areas. Current research trends and directions for future research are also discussed.

Hyper-heuristics are a class of high-level search techniques which operate on a search space of heuristics rather than directly on a search space of solutions. Early hyperheuristics focussed on selecting and applying a low-level heuristic at each stage of a search. Recent trends in hyper-heuristic research have led to a number of approaches being developed to automatically generate new heuristics from a set of heuristic components. This work investigates the suitability of using genetic programming as a hyper-heuristic methodology to generate constructive heuristics to solve the multidimensional 0-1 knapsack problem. A population of heuristics to rank knapsack items are trained on a subset of test problems and then applied to unseen instances. The results over a set of standard benchmarks show that genetic programming can be used to generate constructive heuristics which yield human-competitive results.

Hyper-heuristics are high level search methodologies that operate over a set of heuristics which operate directly on the problem domain. In one of the hyper-heuristic frameworks, the goal is automating the process of selecting a human-designed low level heuristic at each step to construct a solution for a given problem. Constraint Satisfaction Problems (CSP) are well know NP complete problems. In this study, behaviours of two variable ordering heuristics Max-Conflicts (MXC) and Saturation Degree (SD) with respect to various combinations of constraint density and tightness values are investigated in depth over a set of random CSP instances. The empirical results show that the performance of these two heuristics are somewhat complementary and they vary for changing constraint density and tightness value pairs. The outcome is used to design three hyper-heuristics using MXC and SD as low level heuristics to construct a solution for unseen CSP instances. It has been observed that these hyper-heuristics improve the performance of individual low level heuristics even further in terms of mean consistency checks for some CSP instances.

Constraint Satisfaction Problems (CSP) represent an important topic of study because of their many applications in different areas of artificial intelligence and operational research. When solving a CSP, the order in which the variables are selected to be instantiated and the order of the corresponding values to be tried affect the complexity of the search. Hyper-heuristics are flexible methods that provide generality when solving different problems and, within CSP, they can be used to determine the next variable and value to try. They select from a set of low-level heuristics and decide which one to apply at each decision point according to the problem state. This study explores a hyper-heuristic model for variable and value ordering within CSP based on a decision matrix hyper-heuristic that is constructed by going into a local improvement method that changes small portions of the matrix. The results suggest that the approach is able to combine the strengths of different low-level heuristics to perform well on a wide range of instances and compensate for their weaknesses on specific instances.

Hyper-heuristics represent a novel search methodology that is motivated by the goal of automating the process of selecting or combining simpler heuristics in order to solve hard computational search problems. An extension of the original hyper-heuristic idea is to generate new heuristics which are not currently known. These approaches operate on a search space of heuristics rather than directly on a search space of solutions to the underlying problem which is the case with most meta-heuristics implementations. In the majority of hyper-heuristic studies so far, a framework is provided with a set of human designed heuristics, taken from the literature, and with good measures of performance in practice. A less well studied approach aims to generate new heuristics from a set of potential heuristic components. The purpose of this chapter is to discuss this class of hyper-heuristics, in which Genetic Programming is the most widely used methodology. A detailed discussion is presented including the steps needed to apply this technique, some representative case studies, a literature review of related work, and a discussion of relevant issues. Our aim is to convey the exciting potential of this innovative approach for automating the heuristic design process.

In my PhD. dissertation, I focus on action planning and constrained discrete optimization. I try to introduce novel approaches to the field of single-agent planning by combining standard techniques with meta-heuristic optimization, machine-learning algorithms, hyper-heuristics and algorithm selection approaches. Our main goal is to create new and flexible planning algorithms which would be suited for a large variety of real-life problems. Planning is a fundamental and difficult problem in AI and any new results in this area are directly applicable to many other fields. They can be used for single-agent or multi-agent action selection in both competitive or cooperative environment and as we focus on optimization, our techniques are suitable for real-life problems that arise in robotics or transportation.

It is common to find that training of selection hyper-heuristics is done perturbatively. The process usually starts with a random selection module and iterates over a set of instances until finding appropriate values for such module. In this work, however, we present a model for creating selection hyper-heuristics constructively. To achieve so, we use a set of instances evolved for such a task. For each low-level heuristic, we evolved a set of problem instances that are more easily solvable by that particular heuristic (compared to the other ones). Each group contains instances easily solvable with the corresponding heuristic but not with the remaining ones. Thus, our model creates its selector by calculating the centroid of each group. For doing so, the model defines a rule that maps said centroid to one corresponding action (in this case, a low-level heuristic). To test our approach, we select the one-dimensional Bin Packing Problem and set four target performance levels (which we refer to as deltas) for instance generation. Then, we analyze all the possible combinations of deltas. We study how performance of the generated hyper-heuristics shift when they are created using a different number of instances. Our data shows the feasibility of creating a hyper-heuristic under the stated conditions. Effectiveness of the model depends on the deltas used though we observed that higher deltas are useful while lower deltas are not. For example, when considering a delta level of 2.0, our method produced hyper-heuristics with an accumulated average waste 12% lower than that of the best heuristic. But, for a delta level of 0.5, it became impossible to outperform the heuristics.

This article presents a method based on the multi-objective evolutionary algorithm NSGA-II to approximate hyper-heuristics for solving irregular 2D cutting stock problems under multiple objectives. In this case, additionally to the traditional objective of minimizing the number of sheets used to fit a finite number of irregular pieces, the time required to perform the placement task is also minimized, leading to a bi-objective minimization problem with a tradeoff between the number of sheets and the time required for placing all pieces. We solve this problem using multi-objective hyper-heuristics (MOHHs), whose main idea consists of finding a set of simple heuristics which can be combined to find a general solution for a wide range of problems, where a single heuristic is applied depending on the current condition of the problem, instead of applying a unique single heuristic during the whole placement process. The MOHHs are approximated after going through a learning process by mean of the NSGA-II, which evolves combinations of condition-action rules producing at the end a set of Pareto-optimal MOHHs. We tested the approximated MMOHHs on several sets of benchmark problems, having outstanding results for most of the cases.

Hyper-heuristics are methodologies that allow us to selectively apply the most suitable heuristic given the properties of the problem at hand. They can be applied in CSP in different ways, but one way which has received attention in recent years is variable ordering by using hyper-heuristics. To select the next variable, a set of heuristics exist and the hyper-heuristic decides, considering the features that describe the instance at hand, which heuristic is more suitable to be applied at the moment. This paper explores a hyperheuristic model for variable ordering within CSP based on vector hyper-heuristics. Each hyper-heuristic is represented as a set of vectors that maps instance features to heuristics. These vector hyper-heuristics are constructed by going into a local search method that modifies the hyper-heuristics. The results suggest that the approach is able to combine the strengths of different heuristics to perform well on a wide range of instances and compensate for their weaknesses on specific instances, resulting in an improvement in the performance of the search compared against the heuristics applied in isolation.