Jerry Swan

Contrasting meta-learning and hyper-heuristic research: the role of evolutionary algorithms

The fields of machine meta-learning and hyper-heuristic optimisation have developed mostly independently of each other, although evolutionary algorithms (particularly genetic programming) have recently played an important role in the development of both fields. Recent work in both fields shares a common goal, that of automating as much of the algorithm design process as possible. In this paper we first provide a historical perspective on automated algorithm design, and then we discuss similarities and differences between meta-learning in the field of supervised machine learning (classification) and hyper-heuristics in the field of optimisation. This discussion focuses on the dimensions of the problem space, the algorithm space and the performance measure, as well as clarifying important issues related to different levels of automation and generality in both fields. We also discuss important research directions, challenges and foundational issues in meta-learning and hyper-heuristic research. It is important to emphasize that this paper is not a survey, as several surveys on the areas of meta-learning and hyper-heuristics (separately) have been previously published. The main contribution of the paper is to contrast meta-learning and hyper-heuristics methods and concepts, in order to promote awareness and cross-fertilisation of ideas across the (by and large, non-overlapping) different communities of meta-learning and hyper-heuristic researchers. We hope that this cross-fertilisation of ideas can inspire interesting new research in both fields and in the new emerging research area which consists of integrating those fields.

Effective learning hyper-heuristics for the course timetabling problem

Course timetabling is an important and recurring administrative activity in most educational institutions. This article combines a general modeling methodology with effective learning hyper-heuristics to solve this problem. The proposed hyper-heuristics are based on an iterated local search procedure that autonomously combines a set of move operators. Two types of learning for operator selection are contrasted: a static (offline) approach, with a clear distinction between training and execution phases; and a dynamic approach that learns on the fly. The resulting algorithms are tested over the set of real-world instances collected by the first and second International Timetabling competitions. The dynamic scheme statistically outperforms the static counterpart, and produces competitive results when compared to the state-of-the-art, even producing a new best-known solution. Importantly, our study illustrates that algorithms with increased autonomy and generality can outperform human designed problem-specific algorithms.

Searching the Hyper-heuristic Design Space

We extend a previous mathematical formulation of hyper-heuristics to reflect the emerging generalization of the concept. We show that this leads naturally to a recursive definition of hyper-heuristics and to a division of responsibility that is suggestive of a blackboard architecture, in which individual heuristics annotate a shared workspace with information that may also be exploited by other heuristics. Such a framework invites consideration of the kind of relaxations of the domain barrier that can be achieved without loss of generality. We give a concrete example of this architecture with an application to the 3-SAT domain that significantly improves on a related token-ring hyper-heuristic.

The automatic generation of mutation operators for genetic algorithms

We automatically generate mutation operators for Genetic Algorithms (GA) and tune them to problem instances drawn from a given problem class. By so doing, we perform metalearning in which the base-level contains GAs (which learn about problem instances), and the meta-level contains GAmutation operators (which learn about problem classes). We use Register Machines to explore a constrained design space for mutation operators. We show how two commonly used mutation operators (viz. one-point and uniform mutation) can be expressed in this framework. Iterated local search is used to search the space of mutation operators, and on a test-bed of 7 problem classes we identify machine-designed mutation operators which outperform their human counterparts.

HYPERION: a recursive hyper-heuristic framework

Hyper-heuristics are methodologies used to search the space of heuristics for solving computationally difficult problems. We describe an object-oriented domain analysis for hyper-heuristics that orthogonally decomposes the domain into generative policy components. The framework facilitates the recursive instantiation of hyper-heuristics over hyper-heuristics, allowing further exploration of the possibilities implied by the hyper-heuristic concept. We describe Hyperion, a JavaTM class library implementation of this domain analysis.

Pattern-guided genetic programming

Online progress in search and optimization is often hindered by neutrality in the fitness landscape, when many genotypes map to the same fitness value. We propose a method for imposing a gradient on the fitness function of a metaheuristic (in this case, Genetic Programming) via a metric (Minimum Description Length) induced from patterns detected in the trajectory of program execution. These patterns are induced via a decision tree classifier. We apply this method to a range of integer and boolean-valued problems, significantly outperforming the standard approach. The method is conceptually straightforward and applicable to virtually any metaheuristic that can be appropriately instrumented.

Automatically designing selection heuristics

In a standard evolutionary algorithm such as genetic algorithms (GAs), a selection mechanism is used to decide which individuals are to be chosen for subsequent mutation. Examples of selection mechanisms are fitness-proportional selection, in which individuals are chosen with a probability directly in proportion to their fitness value, and rank selection, in which individuals are selected with a probability in proportion to their ordinal ranking by fitness. These two human-designed selection heuristics implicitly assume that fitter individuals produce fitter offspring. Whilst one might invest human ingenuity in the construction of alternative selection heuristics, the approach adopted in this paper is to represent a generic family of selection heuristics which are applied via an algorithmic framework. We then generate instances of selection heuristics and test their performance in an evolutionary algorithm (which in this paper tackles a variety of bitstring optimization problems). The representation we use for the program space is a register machine (a set of real-valued registers on which a program is executed). Fitness-proportional and rank selection can be expressed as one-line programs, and more sophisticated selection heuristics may also be expressed. The result is a system which produces selection heuristics that outperform either of the original selection heuristics.

Automated schematization for web service applications

For the purposes of this paper, a schematic map is a diagrammatic representation based on linear abstractions of networks. With the advent of technologies for web-based delivery of geospatial services it is essential to develop map generalization applications tailored for the same. This paper is concerned with the problem of producing automated schematic maps for web map applications. The paper looks at how previous solutions to the spatial conflict reduction can be adapted and applied to production of automated schematic maps for web services.

Templar – A Framework for Template-Method Hyper-Heuristics

In this work we introduce Templar, a software framework for customising algorithms via the generative technique of template-method hyper-heuristics. We first discuss the need for such an approach, presenting Quicksort as an example. We provide a functional definition of template-method hyper-heuristics, describe how this is implemented by Templar, and show how Templar may be invoked using simple client-code. Finally, we describe experiments using Templar to define a ‘hyper-quicksort’ with the aim of reducing power consumption—the results demonstrate that the generated algorithm has significantly improved performance on the test set.

Object-Oriented Genetic Improvement for Improved Energy Consumption in Google Guava

In this work we use metaheuristic search to improve Google’s Guava library, finding a semantically equivalent version of com.google.common.collect.ImmutableMultimap with reduced energy consumption. Semantics-preserving transformations are found in the source code, using the principle of subtype polymorphism. We introduce a new tool, Opacitor, to deterministically measure the energy consumption, and find that a statistically significant reduction to Guava’s energy consumption is possible. We corroborate these results using Jalen, and evaluate the performance of the metaheuristic search compared to an exhaustive search—finding that the same result is achieved while requiring almost 200 times fewer fitness evaluations. Finally, we compare the metaheuristic search to an independent exhaustive search at each variation point, finding that the metaheuristic has superior performance.