Richard Dazeley

A survey of multi-objective sequential decision-making

Sequential decision-making problems with multiple objectives arise naturally in practice and pose unique challenges for research in decision-theoretic planning and learning, which has largely focused on single-objective settings. This article surveys algorithms designed for sequential decision-making problems with multiple objectives. Though there is a growing body of literature on this subject, little of it makes explicit under what circumstances special methods are needed to solve multi-objective problems. Therefore, we identify three distinct scenarios in which converting such a problem to a single-objective one is impossible, infeasible, or undesirable. Furthermore, we propose a taxonomy that classifies multi-objective methods according to the applicable scenario, the nature of the scalarization function (which projects multi-objective values to scalar ones), and the type of policies considered. We show how these factors determine the nature of an optimal solution, which can be a single policy, a convex hull, or a Pareto front. Using this taxonomy, we survey the literature on multi-objective methods for planning and learning. Finally, we discuss key applications of such methods and outline opportunities for future work.

Empirical evaluation methods for multiobjective reinforcement learning algorithms

While a number of algorithms for multiobjective reinforcement learning have been proposed, and a small number of applications developed, there has been very little rigorous empirical evaluation of the performance and limitations of these algorithms. This paper proposes standard methods for such empirical evaluation, to act as a foundation for future comparative studies. Two classes of multiobjective reinforcement learning algorithms are identified, and appropriate evaluation metrics and methodologies are proposed for each class. A suite of benchmark problems with known Pareto fronts is described, and future extensions and implementations of this benchmark suite are discussed. The utility of the proposed evaluation methods are demonstrated via an empirical comparison of two example learning algorithms.

Authorship Attribution for Twitter in 140 Characters or Less

Authorship attribution is a growing field, moving from beginnings in linguistics to recent advances in text mining. Through this change came an increase in the capability of authorship attribution methods both in their accuracy and the ability to consider more difficult problems. Research into authorship attribution in the 19th century considered it difficult to determine the authorship of a document of fewer than 1000 words. By the 1990s this values had decreased to less than 500 words and in the early 21st century it was considered possible to determine the authorship of a document in 250 words. The need for this ever decreasing limit is exemplified by the trend towards many shorter communications rather than fewer longer communications, such as the move from traditional multi-page handwritten letters to shorter, more focused emails. This trend has also been shown in online crime, where many attacks such as phishing or bullying are performed using very concise language. Cybercrime messages have long been hosted on Internet Relay Chats (IRCs) which have allowed members to hide behind screen names and connect anonymously. More recently, Twitter and other short message based web services have been used as a hosting ground for online crimes. This paper presents some evaluations of current techniques and identifies some new preprocessing methods that can be used to enable authorship to be determined at rates significantly better than chance for documents of 140 characters or less, a format popularised by the micro-blogging website Twitter1. We show that the SCAP methodology performs extremely well on twitter messages and even with restrictions on the types of information allowed, such as the recipient of directed messages, still perform significantly higher than chance. Further to this, we show that 120 tweets per user is an important threshold, at which point adding more tweets per user gives a small but non-significant increase in accuracy.

Consensus clustering and supervised classification for profiling phishing emails in internet commerce security

This article investigates internet commerce security applications of a novel combined method, which uses unsupervised consensus clustering algorithms in combination with supervised classification methods. First, a variety of independent clustering algorithms are applied to a randomized sample of data. Second, several consensus functions and sophisticated algorithms are used to combine these independent clusterings into one final consensus clustering. Third, the consensus clustering of the randomized sample is used as a training set to train several fast supervised classification algorithms. Finally, these fast classification algorithms are used to classify the whole large data set. One of the advantages of this approach is in its ability to facilitate the inclusion of contributions from domain experts in order to adjust the training set created by consensus clustering. We apply this approach to profiling phishing emails selected from a very large data set supplied by the industry partners of the Centre for Informatics and Applied Optimization. Our experiments compare the performance of several classification algorithms incorporated in this scheme.

Automated unsupervised authorship analysis using evidence accumulation clustering

Authorship Analysis aims to extract information about the authorship of documents from features within those documents. Typically, this is performed as a classification task with the aim of identifying the author of a document, given a set of documents of known authorship. Alternatively, unsupervised methods have been developed primarily as visualisation tools to assist the manual discovery of clusters of authorship within a corpus by analysts. However, there is a need in many fields for more sophisticated unsupervised methods to automate the discovery, profiling and organisation of related information through clustering of documents by authorship. An automated and unsupervised methodology for clustering documents by authorship is proposed in this paper. The methodology is named NUANCE, for n -gram Unsupervised Automated Natural Cluster Ensemble. Testing indicates that the derived clusters have a strong correlation to the true authorship of unseen documents.

On the Limitations of Scalarisation for Multi-objective Reinforcement Learning of Pareto Fronts

Multiobjective reinforcement learning (MORL) extends RL to problems with multiple conflicting objectives. This paper argues for designing MORL systems to produce a set of solutions approximating the Pareto front, and shows that the common MORL technique of scalarisation has fundamental limitations when used to find Pareto-optimal policies. The work is supported by the presentation of three new MORL benchmarks with known Pareto fronts.

Recentred local profiles for authorship attribution

Authorship attribution methods aim to determine the author of a document, by using information gathered from a set of documents with known authors. One method of performing this task is to create profiles containing distinctive features known to be used by each author. In this paper, a new method of creating an author or document profile is presented that detects features considered distinctive, compared to normal language usage. This recentreing approach creates more accurate profiles than previous methods, as demonstrated empirically using a known corpus of authorship problems. This method, named recentred local profiles, determines authorship accurately using a simple ‘best matching author’ approach to classification, compared to other methods in the literature. The proposed method is shown to be more stable than related methods as parameter values change. Using a weighted voting scheme, recentred local profiles is shown to outperform other methods in authorship attribution, with an overall accuracy of 69.9% on the ad-hoc authorship attribution competition corpus, representing a significant improvement over related methods.

Constructing Stochastic Mixture Policies for Episodic Multiobjective Reinforcement Learning Tasks

Multiobjective reinforcement learning algorithms extend reinforcement learning techniques to problems with multiple conflicting objectives. This paper discusses the advantages gained from applying stochastic policies to multiobjective tasks and examines a particular form of stochastic policy known as a mixture policy. Two methods are proposed for deriving mixture policies for episodic multiobjective tasks from deterministic base policies found via scalarised reinforcement learning. It is shown that these approaches are an efficient means of identifying solutions which offer a superior match to the users preferences than can be achieved by methods based strictly on deterministic policies.

Automatically determining phishing campaigns using the USCAP methodology

Phishing fraudsters attempt to create an environment which looks and feels like a legitimate institution, while at the same time attempting to bypass filters and suspicions of their targets. This is a difficult compromise for the phishers and presents a weakness in the process of conducting this fraud. In this research, a methodology is presented that looks at the differences that occur between phishing websites from an authorship analysis perspective and is able to determine different phishing campaigns undertaken by phishing groups. The methodology is named USCAP, for Unsupervised SCAP, which builds on the SCAP methodology from supervised authorship and extends it for unsupervised learning problems. The phishing website source code is examined to generate a model that gives the size and scope of each of the recognized phishing campaigns. The USCAP methodology introduces the first time that phishing websites have been clustered by campaign in an automatic and reliable way, compared to previous methods which relied on costly expert analysis of phishing websites. Evaluation of these clusters indicates that each cluster is strongly consistent with a high stability and reliability when analyzed using new information about the attacks, such as the dates that the attack occurred on. The clusters found are indicative of different phishing campaigns, presenting a step towards an automated phishing authorship analysis methodology.

Unsupervised authorship analysis of phishing webpages

Authorship analysis on phishing websites enables the investigation of phishing attacks, beyond basic analysis. In authorship analysis, salient features from documents are used to determine properties about the author, such as which of a set of candidate authors wrote a given document. In unsupervised authorship analysis, the aim is to group documents such that all documents by one author are grouped together. Applying this to cyber-attacks shows the size and scope of attacks from specific groups. This in turn allows investigators to focus their attention on specific attacking groups rather than trying to profile multiple independent attackers. In this paper, we analyse phishing websites using the current state of the art unsupervised authorship analysis method, called NUANCE. The results indicate that the application produces clusters which correlate strongly to authorship, evaluated using expert knowledge and external information as well as showing an improvement over a previous approach with known flaws.