Adaptive Markov Chain Monte Carlo (MCMC) adapts the covariance of the proposal distribution to improve the efficiency of Metropolis Hastings (MH). Adaptive Metropolis (AM) is the prime example. Some stochastic optimisation techniques adapt the co-variance of the search distribution. Some examples are Gaussian Adaptation (GaA) and (1+1)-Covariance Matrix Adaptation Evolution Strategy (CMAES) that can be turned into MCMC samplers in a straightforward way. However, the adaptation rational used by these samplers differ. AM estimates the covariance of the target distribution based on the generated samples. GaA adapts the covariance such that the entropy of the proposal is increased/decreased when the candidate sample is accepted/rejected while adaptation in CMAES increases the likelihood of generating better search points. We compare the performance of AM, GaA and (1+1)-CMAES samplers on a test suite of target distributions to understand the effectiveness of the adaptation mechanism used.

We evaluate two (1+1)-natural evolution strategies (NES) turned into adaptive Markov chain Monte Carlo (MCMC) samplers on a test suite of probability distributions. We compare their performance with the AM-family of samplers considered to be the state of the art in adaptive MCMC. Our experiments show that natural gradient based adaptation used in NES further improves adaptive MCMC.

Machine learning for data mining applications in the field of bioinformatics is to extract new knowledge to provide an improved and effective diagnosis process for patients. In this paper, we introduce an adaptive ensemble learning for classifying high-dimensional multi-class imbalanced genomic data. The aspect is to design and develop an optimal ensemble method for information discovery on genomic data, which improve the prediction accuracy of DNA variant classification. The proposed method is based on ensemble of decision trees, data pre-processing, feature selection and grouping. It converts an imbalanced genomic data into multiple balanced ones and then builds a number of decision trees on these multiple data with specific feature groups. The outputs of these trees are combined for classifying new instances by majority voting technique. In this empirical study, different ensemble predictive modelling techniques like Random Forest, Boosting and Bagging were compared with the proposed ensemble method. The experimental results on genomic data (148 Exome datasets) of Brugada syndrome from the Centre of Medical Genetics, VUB UZ Brussel show that the proposed method is usually superior to the conventional ensemble learning algorithms when classifying the high-dimensional multi-class imbalanced genomic data.

The exome or genome based high throughput screening techniques are becoming a definitive criterion in the conventional clinical analysis of the genetic diseases. However, pathogenic classification of an identified variant, is still a manual and time consuming process for clinical geneticists. Thus, to facilitate the variant classification process, we have developed GeVaCT, a Java based tool that implements a classification approach based on the literature review of cardiac arrhythmia syndromes. Furthermore, the adoption of this automated knowledge engineer by the clinical geneticists will aid to build a knowledge base for the evolution of the variant classification process by use of novel machine learning approaches.

We turn the Covariance Matrix Adaptation Evolution Strategy into an adaptive Markov Chain Monte Carlo (or MCMC) sampling algorithm that adapts online to the target distribution, i.e. the distribution to be sampled from. We call the resulting algorithm CMA-Sampling. It exhibits a higher convergence rate, a better mixing, and consequently a more effective MCMC sampler. We look at a few variants and compare their adaptiveness to a number of other adaptive samplers, including Haario et. al's AM sampler, on a testsuite of 4 target distributions.

In this paper, we introduce a new adaptive rule-based classifier for multi-class classification of biological data, where several problems of classifying biological data are addressed: overfitting, noisy instances and class-imbalance data. It is well known that rules are interesting way for representing data in a human interpretable way. The proposed rule-based classifier combines the random subspace and boosting approaches with ensemble of decision trees to construct a set of classification rules without involving global optimisation. The classifier considers random subspace approach to avoid overfitting, boosting approach for classifying noisy instances and ensemble of decision trees to deal with class-imbalance problem. The classifier uses two popular classification techniques: decision tree and k-nearest-neighbor algorithms. Decision trees are used for evolving classification rules from the training data, while k-nearest-neighbor is used for analysing the misclassified instances and removing vagueness between the contradictory rules. It considers a series of k iterations to develop a set of classification rules from the training data and pays more attention to the misclassified instances in the next iteration by giving it a boosting flavour. This paper particularly focuses to come up with an optimal ensemble classifier that will help for improving the prediction accuracy of DNA variant identification and classification task. The performance of proposed classifier is tested with compared to well-approved existing machine learning and data mining algorithms on genomic data (148 Exome data sets) of Brugada syndrome and 10 real benchmark life sciences data sets from the UCI (University of California, Irvine) machine learning repository. The experimental results indicate that the proposed classifier has exemplary classification accuracy on different types of biological data. Overall, the proposed classifier offers good prediction accuracy to new DNA variants classification where noisy and misclassified variants are optimised to increase test performance. (C) 2016 Elsevier Ltd. All rights reserved.

Dynamic webservice composition is a promising ICT support service for virtual organizations.However, dynamic webservice composition remains a nondeterministic polynomial (NP) hard problem despite more than 10 years of extensive research, making the applicability of the technique to problems of industrial relevance limited.In [48], we proposed a layered method, SLUM, to combat the problem.Analytically, SLUM overcomes the relative weaknesses of two widely used approaches in the literature -the local planning (hereafter L-MIP) strategy and the Mixed Integer Programming (S-MIP) method.Despite the promising benefits of SLUM, it's unknown to what extent and under what circumstances SLUM is better or worse than L-MIP algorithms and S-MIP.The research objective of the study was to investigate the relative performance of SLUM w.r.t S-MIP and L-MIP using two performance criteria: -solution quality and CPU running time.Several randomly generated two task workflows of monotonically increasing hardness in the number of webservices per task were used to benchmark SLUM against the other two algorithms.A set of numerical and statistical techniques were used to experimentally compare the solution quality and the running time growth of SLUM against L-MIP and S-MIP.We determined that SLUM generates solutions with an average quality of 93% w.r.t the global optimum.Further, we show that SLUM yields solutions that are 5% more quality than L-MIP.On the other hand, we established that L-MIP outperforms both S-MIP and SLUM by multiple factors in terms of computational efficiency.However, we find that for problem instances with less than 22 webservices per task, S-MIP is about 1.3 times faster than SLUM.Beyond n=22, the running time of SLUM t eB , expressed in terms of the running time of S-MIP t eA , is given by t eB = t eA 0.78 .We also establish that SLUM is asymptotically 3.6 times faster than S-MIP on average.We conclude that in order for a virtual enterprise broker to obtain maximum benefit from dynamic service composition, the broker should combine the three techniques in the following manner-(1) for service request without global constraints requirements, L-MIP is the most suitable method to use, (2) Where there is need for global constraints and the number of service providers per task is less than 22, S-MIP is most preferred and (3) in scenarios the number of service providers per task is more than 22 and there is a need to satisfy global constraints, SLUM is superior to both S-MIP and L-MIP.

In this paper, we propose the use of schemata bandits for optimisation. This technique is a subclass of hierarchical bandits where the bandits are schemata. We investigate its use on a benchmark of binary combinatorial optimization problems, the Maximum Satisfiability (MAXSAT) problem. We compare performance with hierarchical Bayesian Optimization Algorithms (hBOAs), namely GSAT and WALKSAT. Results suggest that using a bandit strategy enhances solver performance.

High-dimensional genomic big data with hundred of features present a big challenge in cluster analysis. Usually, genomic data are noisy and have correlation among the features. Also, different subspaces exist in high-dimensional genomic data. This paper presents a feature selecting and grouping method for ensemble clustering of high-dimensional genomic data. Two most popular clustering methods: k-means and similarity-based clustering are used for ensemble clustering. Ensemble clustering is more effective in clustering high-dimensional complex data than the traditional clustering algorithms. In this paper, we cluster un-labeled genomic data (148 Exome data sets) of Brugada syndrome from the Centre of Medical Genetics, VUB UZ Brussel using SimpleKMeans, XMeans, DBScan, and MakeDensityBasedCluster algorithms and compare the clustering results with proposed ensemble clustering method. Furthermore, we use biclustering (δ-Biclustering) algorithm on each cluster to find the sub-matrices in the genomic data, which clusters both instances and features simultaneously.

Clustering of high-dimensional biological big data is incredibly difficult and challenging task, as the data space is often too big and too messy.The conventional clustering methods can be inefficient and ineffective on high-dimensional biological big data, because traditional distance measures may be dominated by the noise in many dimensions.An additional challenge in biological big data is that we need to find not only the clusters of instances (genes), but also for each cluster a set of features (conditions) that manifest the cluster.In this paper, we propose an ensemble clustering approach with feature selection and grouping for clustering high-dimensional biological big data.It uses two well-approved clustering methods: (a) k-means clustering and (b) similarity-based clustering.This approach selects the most relevant features in the dataset and grouping them into subset of features to overcome the problems associated with the traditional clustering methods.Also, we applied biclustering on each cluster that generated by ensemble clustering to find the sub-matrices in the biological data by the mean squared residue scores.We have applied the proposed clustering method on unlabeled genomic data (148 Exome datasets) of Brugada syndrome to discover previously unknown data patterns.Experiments verify that the proposed clustering method achieved high performance clustering results on high-dimensional biological big data.

Abstract Biologists have mostly studied under what circumstances honest signaling is stable. Stability, however, is not sufficient to explain the emergence of honest signaling. We study the evolution of honest signaling between selfish, adaptive individuals and observe that honest signaling can emerge through learning. More importantly, honest signaling may emerge in cases where it is not evolutionary stable. In such cases, honesty and dishonesty co-exist. Furthermore, honest signaling does not necessarily emerge in cases where it is evolutionary stable. We show that the latter is due to the existence of other, more important equilibria and that the importance of equilibria is related to Pareto-optimality.

The most interesting challenge for a reinforcement learning agent is to learn online in unknown large discrete, or continuous stochastic model. The agent has not only to trade-off between exploration and exploitation, but also has to find a good set of basis functions to approximate the value function. We extend offline kernel-based LSPI (or least squares policy iteration) to online learning. Online kernel-based LSPI combines feature of offline kernel-based LSPI and online LSPI. Online kernel-based LSPI uses knowledge gradient policy as an exploration policy to trade-off between exploration and exploitation, and the approximate linear dependency based kernel sparsification method to select basis functions automatically. We compare between online kernel-based LSPI and online LSPI on 5 discrete Markov decision problems, where online kernel-based LSPI outperforms online LSPI according to the optimal policy performance.

Student retention has become one of the most important priorities for decision makers in higher learning institutions (HLI). Improving student retention starts with a thorough understanding of the reasons behind the attrition. In this study, using student demographic and institutional data along with several business intelligence (BI) techniques, we developed prototype to predict likelihood of student persistence or dropout. This study used classification models generated using Waikato Environment for Knowledge Analysis (WEKA). The model was built using the 10-fold cross validation, and holdout method (60% of the data was used as training and the remaining as test and validation). Random sampling techniques were used in selecting the datasets. The attribute selection analysis of the models revealed that the student age on entry, parent occupation, health of student and financial variables are among the most important predictors of the phenomenon. Results of the classifiers were compared using accuracy level, confusion matrices and speed of model building benchmarks. The study shows that identifying the relevant student background factors can be incorporated to design a prototype that can serve as valuable tool in predicting student withdrawal or persistence as well as recommend the necessary intervention strategies to adopt, leading to better education efficiency. Keywords-Business Intelligence, Retention, attrition, WEKA, classifiers.

Medical literature have recognized physical activity as a key factor for a healthy life due to its remarkable benefits. However, there is a great variety of physical activities and not all of them have the same effects on health nor require the same effort. As a result, and due to the ubiquity of commodity devices able to track users' motion, there is an increasing interest on performing activity recognition in order to detect the type of activity carried out by the subjects and being able to credit them for their effort, which has been detected as a key requirement to promote physical activity. This paper proposes a novel approach for performing activity recognition using Monte Carlo Schemata Search (MCSS) for feature selection and random forests for classification. To validate this approach we have carried out an evaluation over PAMAP2, a public dataset on physical activity available in UCI Machine Learning repository, enabling replication and assessment. The experiments are conducted using leave-one-subject-out cross validation and attain classification accuracies of over 93% by using roughly one third of the total set of features. Results are promising, as they outperform those obtained in other works on the same dataset and significantly reduce the set of features used, which could translate in a decrease of the number of sensors required to perform activity recognition and, as a result, a reduction of costs.

A multi-objective multi-armed bandit (MOMAB) problem is a sequential decision process with stochastic reward vectors. We extend knowledge gradient (KG) policy to the MOMAB problem, and we propose Pareto-KG and scalarized-KG algorithms. The Pareto-KG trades off between exploration and exploitation by combining KG policy with Pareto dominance relations. The scalarized-KG makes use of a linear or non-linear scalarization function to convert the MOMAB problem into a single-objective multi-armed bandit problem and uses KG policy to trade off between exploration and exploitation. To measure the performance of the proposed algorithms, we introduce three regret measures. We compare empirically the performance of the KG policy with UCB1 policy on a test suite of MOMAB problems with normal distributions. The Pareto-KG and scalarized-KG are the algorithms with the best empirical performance.

The multi-objective multi-armed bandit (MOMAB) problem is a se- quential decision process with stochastic rewards. Each arm generates a vector of rewards instead of a single scalar reward. Moreover, these multiple rewards might be conflicting. The MOMAB-problem has a set of Pareto optimal arms and an agent's goal is not only to find that set but also to play evenly or fairly the arms in that set. To find the Pareto optimal arms, linear scalarized function or Pareto dominance relations can be used. The linear scalarized function converts the multi- objective optimization problem into a single objective one and is a very popular ap- proach because of its simplicity. The Pareto dominance relations optimizes directly the multi-objective problem. In this paper, we extend the Thompson Sampling pol- icy to be used in theMOMAB problem. We propose Pareto Thompson Sampling and linear scalarized Thompson Sampling approaches. We compare empirically between Pareto Thompson Sampling and linear scalarized Thompson Sampling on a test suite ofMOMAB problems with Bernoulli distributions. Pareto Thompson Sampling is the approach with the best empirical performance.

In the stochastic multi-objective multi-armed bandit (MOMAB), arms generate a vector of stochastic normal rewards, one per objective, instead of a single scalar reward. As a result, there is not only one optimal arm, but there is a set of optimal arms (Pareto front) using Pareto dominance relation. The goal of an agent is to find the Pareto front. To find the optimal arms, the agent can use linear scalarization function that transforms a multi-objective problem into a single problem by summing the weighted objectives. Selecting the weights is crucial, since different weights will result in selecting a different optimum arm from the Pareto front. Usually, a predefined weights set is used and this can be computational inefficient when different weights will optimize the same Pareto optimal arm and arms in the Pareto front are not identified. In this paper, we propose a number of techniques that adapt the weights on the fly in order to ameliorate the performance of the scalarized MOMAB. We use genetic and adaptive scalarization functions from multi-objective optimization to generate new weights. We propose to use Thompson sampling policy to select frequently the weights that identify new arms on the Pareto front. We experimentally show that Thompson sampling improves the performance of the genetic and adaptive scalarization functions. All the proposed techniques improves the performance of the standard scalarized MOMAB with a fixed set of weights.

An empirical comparative study is made of a sample of action selection policies on a test suite of the Bernoulli multi-armed bandit with K = 10, K = 20 and K = 50 arms, each for which we consider several success probabilities. For such problems the rewards are either Success or Failure with unknown success rate. Our study focusses on - greedy, UCB1-Tuned, Thompson sampling, the Gittin's index policy, the knowledge gradient and a new hybrid algorithm. The last two are not well- known in computer science. In this paper, we examine policy dependence on the horizon and report results which suggest that a new hybridized procedure based on Thompsons sampling improves on its regret.