Shahin Rostami

Covariance matrix adaptation pareto archived evolution strategy with hypervolume-sorted adaptive grid algorithm

Real-world problems often involve the optimisation of multiple conflicting objectives. These problems, referred to as multi-objective optimisation problems, are especially challenging when more than three objectives are considered simultaneously. This paper proposes an algorithm to address this class of problems. The proposed algorithm is an evolutionary algorithm based on an evolution strategy framework, and more specifically, on the Covariance Matrix Adaptation Pareto Archived Evolution Strategy (CMA-PAES). A novel selection mechanism is introduced and integrated within the framework. This selection mechanism makes use of an adaptive grid to perform a local approximation of the hypervolume indicator which is then used as a selection criterion. The proposed implementation, named Covariance Matrix Adaptation Pareto Archived Evolution Strategy with Hypervolume-sorted Adaptive Grid Algorithm (CMA-PAES-HAGA), overcomes the limitation of CMA-PAES in handling more than two objectives and displays a remarkably good performance on a scalable test suite in five, seven, and ten-objective problems. The performance of CMA-PAES-HAGA has been compared with that of a competition winning meta-heuristic, representing the state-of-the-art in this sub-field of multi-objective optimisation. The proposed algorithm has been tested in a seven-objective real-world application, i.e. the design of an aircraft lateral control system. In this optimisation problem, CMA-PAES-HAGA greatly outperformed its competitors.

A novel preference articulation operator for the Evolutionary Multi-Objective Optimisation of classifiers in concealed weapons detection

The incorporation of decision maker preferences is often neglected in the Evolutionary Multi-Objective Optimisation (EMO) literature. The majority of the research in the field and the development of EMO algorithms is primarily focussed on converging to a Pareto optimal approximation close to or along the true Pareto front of synthetic test problems. However, when EMO is applied to real-world optimisation problems there is often a decision maker who is only interested in a portion of the Pareto front (the Region of Interest) which is defined by their expressed preferences for the problem objectives. In this paper a novel preference articulation operator for EMO algorithms is introduced (named the Weighted Z-score Preference Articulation Operator) with the flexibility of being incorporated a priori, a posteriori or progressively, and as either a primary or auxiliary fitness operator. The Weighted Z-score Preference Articulation Operator is incorporated into an implementation of the Multi-Objective Evolutionary Algorithm Based on Decomposition (named WZ-MOEA/D) and benchmarked against MOEA/D-DRA on a number of bi-objective and five-objective test problems with test cases containing preference information. After promising results are obtained when comparing WZ-MOEA/D to MOEA/D-DRA in the presence of decision maker preferences, WZ-MOEA/D is successfully applied to a real-world optimisation problem to optimise a classifier for concealed weapon detection, producing better results than previously published classifier implementations.

A fast hypervolume driven selection mechanism for many-objective optimisation problems

Abstract Solutions to real-world problems often require the simultaneous optimisation of multiple conflicting objectives. In the presence of four or more objectives, the problem is referred to as a “many-objective optimisation problem”. A problem of this category introduces many challenges, one of which is the effective and efficient selection of optimal solutions. The hypervolume indicator (or s-metric), i.e. the size of dominated objective space, is an effective selection criterion for many-objective optimisation. The indicator is used to measure the quality of a non-dominated set, and can be used to sort solutions for selection as part of the contributing hypervolume indicator. However, hypervolume based selection methods can have a very high, if not infeasible, computational cost. The present study proposes a novel hypervolume driven selection mechanism for many-objective problems, whilst maintaining a feasible computational cost. This approach, named the Hypervolume Adaptive Grid Algorithm (HAGA), uses two-phases (narrow and broad) to prevent population-wide calculation of the contributing hypervolume indicator. Instead, HAGA only calculates the contributing hypervolume indicator for grid populations, i.e. for a few solutions, which are close in proximity (in the objective space) to a candidate solution when in competition for survival. The result is a trade-off between complete accuracy in selecting the fittest individuals in regards to hypervolume quality, and a feasible computational time in many-objective space. The real-world efficiency of the proposed selection mechanism is demonstrated within the optimisation of a classifier for concealed weapon detection.

Progressive Preference Articulation for Decision Making in Multi-Objective Optimisation Problems

This paper proposes a novel algorithm for addressing multi-objective optimisation problems, by employing a progressive preference articulation approach to decision making. This enables the interactive incorporation of problem knowledge and decision maker preferences during the optimisation process. A novel progressive preference articulation mechanism, derived from a statistical technique, is herein proposed and implemented within a multi-objective framework based on evolution strategy search and hypervolume indicator selection. The proposed algorithm is named the Weighted Z-score Covariance Matrix Adaptation Pareto Archived Evolution Strategy with Hypervolume-sorted Adaptive Grid Algorithm (WZ-HAGA). WZ-HAGA is based on a framework that makes use of evolution strategy logic with covariance matrix adaptation to perturb the solutions, and a hypervolume indicator driven algorithm to select successful solutions for the subsequent generation. In order to guide the search towards interesting regions, a preference articulation procedure composed of four phases and based on the weighted z-score approach is employed. The latter procedure cascades into the hypervolume driven algorithm to perform the selection of the solutions at each generation. Numerical results against five modern algorithms representing the state-of-the-art in multi-objective optimisation demonstrate that the proposed WZ-HAGA outperforms its competitors in terms of both the hypervolume indicator and pertinence to the regions of interest.

CMA-PAES: Pareto archived evolution strategy using covariance matrix adaptation for Multi-Objective Optimisation

The quality of Evolutionary Multi-Objective Optimisation (EMO) approximation sets can be measured by their proximity, diversity and pertinence. In this paper we introduce a modular and extensible Multi-Objective Evolutionary Algorithm (MOEA) capable of converging to the Pareto-optimal front in a minimal number of function evaluations and producing a diverse approximation set. This algorithm, called the Covariance Matrix Adaptation Pareto Archived Evolution Strategy (CMA-PAES), is a form of (μ + λ) Evolution Strategy which uses an online archive of previously found Pareto-optimal solutions (maintained by a bounded Pareto-archiving scheme) as well as a population of solutions which are subjected to variation using Covariance Matrix Adaptation. The performance of CMA-PAES is compared to NSGA-II (currently considered the benchmark MOEA in the literature) on the ZDT test suite of bi-objective optimisation problems and the significance of the results are analysed using randomisation testing.

Knowing the unknown: visualising consumption blind-spots in recommender systems

In this paper we consider how to help users to better understand their consumption profiles by examining two approaches to visualising user profiles - chord diagrams, and bar charts - aimed at revealing to users those regions of the recommendation space that are unknown to them, i.e. blind-spots. Both visualisations do this by connecting profile preferences with a filtered recommendation space. We compare and contrast the two visualisations in a live user study (n = 70). The results suggest that, although users can understand both visualisations, chord diagrams are particularly effective in helping users to identify blind-spots, while simpler bar charts are better for conveying what was already known in a profile. Evaluating the understandability of blind-spot visualizations is a first step toward using visual explanations to help address a criticism of recommender systems: that personalising information creates filter bubbles.

Age Detection Through Keystroke Dynamics from User Authentication Failures

In this paper an incident response approach is proposed for handling detections of authentication failures in systems that employ dynamic biometric authentication and more specifically keystroke user recognition. The main component of the approach is a multi layer perceptron focusing on the age classification of a user. Empirical findings show that the classifier can detect the age of the subject with a probability that is far from the uniform random distribution, making the proposed method suitable for providing supporting yet circumstantial evidence during e-discovery.

A multi objective approach to evolving artificial neural networks for coronary heart disease classification

The optimisation of the accuracy of classifiers in pattern recognition is a complex problem that is often poorly understood. Whilst numerous techniques exist for the optimisation of weights in artificial neural networks (e.g. the Widrow-Hoff least mean squares algorithm and back propagation techniques), there do not exist any hard and fast rules for choosing the structure of an artificial neural network - in particular for choosing both the number of the hidden layers used in the network and the size (in terms of number of neurons) of those hidden layers. However, this internal structure is one of the key factors in determining the accuracy of the classification. This paper proposes taking a multi-objective approach to the evolutionary design of artificial neural networks using a powerful optimiser based around the state-of-the-art MOEA/D-DRA algorithm and a novel method of incorporating decision maker preferences. In contrast to previous approaches, the novel approach outlined in this paper allows the intuitive consideration of trade-offs between classification objectives that are frequently present in complex classification problems but are often ignored. The effectiveness of the proposed multi-objective approach to evolving artificial neural networks is then shown on a real-world medical classification problem frequently used to benchmark classification methods.

On the Integrity of Performance Comparison for Evolutionary Multi-objective Optimisation Algorithms.

This paper proposes the notion that the experimental results and performance analyses of newly developed algorithms in the field of multi-objective optimisation may not offer sufficient integrity for hypothesis testing. The reason for this is that many implementations exist of the same optimisation algorithms, and these may vary in behaviour due to the interpretation of the developer. This is demonstrated through the comparison of three implementations of the popular Non-dominated Sorting Genetic Algorithm II (NSGA-II) from well-regarded frameworks using the hypervolume indicator. The results show that of the thirty considered comparison cases, only four indicate that there was no significant difference between the performance of either implementation.

A Comparison of Re-sampling Techniques for Pattern Classification in Imbalanced Data-Sets

Class imbalance is a common challenge when dealing with pattern classification of real-world medical data-sets. An effective counter-measure typically used is a method known as re-sampling. In this paper we implement an ANN with different re-sampling techniques to subsequently compare and evaluate the performances. Re-sampling strategies included a control, under-sampling, over-sampling, and a combination of the two. We found that over-sampling and the combination of under- and over-sampling both led to a significantly superior classifier performance compared to under-sampling only in correctly predicting labelled classes.