No description available.

This thesis presents results from the developemnt, investigation, testing and evaluation of novel meta-heuristic techniques aiming to further improve the <i>state-of-the-art </i>of algorithms for local minima free Neural Network supervised learning. Several approaches for solving Global Optimisation problems that make use of novel meta-heuristic techniques, so called Low-discrepancy Sequences, and hybrid Evolutionary Algorithms are proposed here, investigated and critically discussed. Furthermore, the novel methods are tested on a number of multimodal mathematical function optimisation problems, as well as on a variety of Neural Network learning tasks, including real-world benchmark datasets. Comparison of the results from the investigated methods with such from standard Backpropagation, Evolutionary Algorithms, and other stochastic approaches (Simulated Annealing, Tabu Search, etc.) is conducted in order to demonstrate their competitiveness in terms of number of function evaluations, learning speed and Neural Network generalisation abilities. Finally, the investigated techniques are applied and tested on real-world problems for the intelligent recognition and classification of cork tiles. An Intelligent Computer Vision system is built. The system includes the following stages: image acquisition; image processing (feature extraction and statistical data processing); Neural Network architecture design; supervised learning utilising the proposed Global Optimisation techniques; and finally, extensive system evaluation. The presented examples and case studies demonstrate that the proposed techniques can be effectively applied for the optimisation of mathematical multimodal functions. The investigated methods are successful in local minima free Neural Network learning, and they can be used for solving real-world industrial problems.

Learning Classifier Systems (LCS) traditionally use a ternary encoding to generalise over the environmental inputs and to associate appropriate actions. However, a number of schemes have been presented beyond this, ranging from integers to artificial neural networks. This thesis investigates the use of Dynamical Genetic Programming (DGP) as a knowledge representation within LCS. DGP is a temporally dynamic, graph-based, symbolic representation. Temporal dynamism has been identified as an important aspect in biological systems, artificial life, and cognition in general. Furthermore, discrete dynamical systems have been found to exhibit inherent content-addressable memory. In this thesis, the collective emergent behaviour of ensembles of such dynamical function networks are herein shown to be exploitable toward solving various computational tasks. Significantly, it is shown possible to exploit the variable-length, adaptive memory existing inherently within the networks under an asynchronous scheme, and where all new parameters introduced are self-adaptive. It is shown possible to exploit the collective mechanics to solve both discrete and continuous-valued reinforcement learning problems, and to perform symbolic regression. In particular, the representation is shown to provide improved performance beyond a traditional Genetic Programming benchmark on a number of a composite polynomial regression tasks. Superior performance to previously published techniques is also shown in a continuous-input-output reinforcement learning problem. Finally, it is shown possible to perform multi-step-ahead predictions of a financial time-series by repeatedly sampling the network states at succeeding temporal intervals.

This thesis is concerned with the application of simulated evolution (SE) to the steady-state parameter identification problem of a simulated and real 3-phase induction machine, over the no-load direct-on-line start period. In the case of the simulated 3-phase induction machine, the Kron's two-axis dynamic mathematical model was used to generate the real and simulated system responses where the induction machine parameters remain constant over the entire range of slip. The model was used in the actual value as well as the per-unit system, and the parameters were estimated using both the genetic algorithm (GA) and the evolutionary programming (EP) from the machine's dynamic response to a direct-on-line start. Two measurement vectors represented the dynamic responses and all the parameter identification processes were subject to five different levels of measurement noise. For the case of the real 3-phase induction machine, the real system responses were generated by the real 3-phase induction machine whilst the simulated system responses were generated by the Kron's model. However, the real induction machine's parameters are not constant over the range of slip, because of the nonlinearities caused by the skin effect and saturation. Therefore, the parameter identification of a real3-phase induction machine, using EP from the machine's dynamic response to a direct-on-line start, was not possible by applying the same methodology used for estimating the parameters of the simulated, constant parameters, 3-phase induction machine.

Το αντικείμενο της μεταπτυχιακής αυτής εργασίας είναι η αυτόματη κα-τηγοριο-ποίηση κειμένου (text classification) χρησιμοποιώντας τεχνι-κές μηχανικής μάθη-σης. Με τον όρο κατηγοριοποίηση κειμένου εννοούμε την διαδικασία αυτόματης κατάταξης κειμένων φυσικής γλώσσας σε προκα-θορισμένο αριθμό θεματικών κατηγοριών. Σήμερα, η κατηγοριοποίηση κει-μένου χρησιμοποιείται σε διάφορα περιβάλλοντα εφαρμογών, όπως για πα-ράδειγμα στη δημιουργία ευρετηρίων που προέρχονται από κείμενα, στην αυτόματη κατηγοριοποίηση ειδήσεων, στην κωδικοποίηση βιβλίων σε βι-βλιοθήκες, στο φιλτράρισμα της ηλεκτρονικής αλληλογραφίας (spam emails), στα αποτελέσματα μηχανών αναζήτησης στο διαδίκτυο (π.χ. Yahoo, Google), κ. α. Συνοπτικά, στην παρούσα εργασία: 1. Παρουσιάζεται η σημερινή δραστηριότητα της μηχανικής μάθησης στους τομείς της αυτόματης κατηγοριοποίησης κειμένου. 2. Δείχνεται πειραματικά η υψηλή απόδοση που επιτυγχάνεται με τη χρήση τεχνικών μηχανικής μάθησης για την αντιμετώπιση του προβλήμα-τος. 3. Παρουσιάζεται η δημιουργία ενός νέου αλγόριθμου για την βελτιστοποίηση της ακρίβειας. 4. Ερμηνεύονται τα αποτελέσματα των πραγματοποιηθέντων πειραμάτων και τέλος 5. Γενικεύονται όπου είναι δυνατόν τα συμπεράσματα που προκύπτουν από την ως άνω μελέτη. / -

Recent advances in molecular biology have led to the accumulation of large amounts of data on Protein-Protein Interaction (PPI) networks in different species, such as yeast and humans. Due to the inherent complexity, analysing such volumes of data to extract knowledge, such as protein complexes or regulatory pathways, represents not only an enormous challenge but also a great opportunity. This Thesis explores the application of machine learning approaches to detecting protein complexes from PPI networks obtained by Tandem Affinity Purification/Mass Spectrometry (TAP-MS) experiments. TAP-MS PPI networks are usually constructed as binary, and the co-complex relations are largely ignored. In order to take into account the non-binary information of co-complex relations in T AP-MS PPI networks, a new framework for detecting protein complexes has been proposed. Under this framework, two types of graph clustering algorithms and an integrative evaluation platform combining data-driven and knowledge-based quality measures have been proposed and studied. One type of the proposed graph clustering algorithms is random walk based graph clustering, resulting in Enhanced Random Walk with Restart (ERWR) and Random Walk with Restarting Baits (RWRB). The other type is based on the modelling of TAP-MS PPI networks as bipartite graphs, resulting in the Bipartite Graph based Clustering Algorithm (BGCA). The ER WR algorithm has been developed from the Random Walk with Restart (R WR). The key contribution of the ERWR is the introduction of a tuning factor into the random walk process. The tuning factor strengthens connections between nodes that are closer and weakens those that are distant, so that the random walker prefers moving to nodes which are potentially in the same clusters with the starting node.

This work investigates combinations of two different nature-inspired machine learning algorithms - Evolutionary Algorithms and Temporal Difference Learning. Both algorithms are introduced along with a survey of previous work in the field. A variety of ways of hybridising the two algorithms are considered, falling into two main categories - those where both algorithms operate on the same set of parameters, and those where evolution searches for beneficial parameters to aid Temporal Difference Learning. These potential approaches to hybridisation are explored by applying them to three different problem domains, all loosely linked by the theme of games. The Mountain Car task is a common reinforcement learning benchmark that has been shown to be potentially problematic for neural networks. Ms. Pac-Man is a classic arcade game with a complex virtual environment, and Othello is a popular two-player zero sum board game. Results show that simple hybridisation approaches often do not improve performance, which can be dependent on many factors of the individual algorithms. However, results have also shown that these factors can be successfully tuned by evolution. The main contributions of this thesis are an analysis of the factors that can affect individual algorithm performance, and demonstration of some novel approaches to hybridisation. These consist of use of Evolution Strategies to tune Temporal Difference Learning parameters on multiple problem domains, and evolution of n-tuple configurations for Othello board evaluation. In the latter case, a level of performance was achieved that was competitive with the state of the art.

The idea of creating agents that automatically learn to play games through experience has been one of the major goals for game researchers. Simulation-based search and reinforcement learning have been two of the most active areas of research tackling this problem. One of the main challenges that links both areas is how to acquire domain knowledge that can. be effectively integrated into simulation-based search algorithms. In this thesis we address this challenge in several ways. First, we use temporal difference learning to find value functions in the form of weighted piece counters and N-tuple systems to play the game of Othello. Next, we present an algorithm that combines TD learning with coevolution to learn value functions of higher quality. These learned value functions Serve as basis to enhance the performance of Monte Carlo Tree Search by incorporating them in the tree and default policies. Finally, we conduct an extensive empirical analysis of Monte Carlo Tree Search by comparing it against other simulation-based and minimax search algorithms.

This thesis alms to investigate how machine learning and statistical approaches can be employed to support the analysis of gait patterns captured by wearable sensors data. The thesis has proposed a machine learning and statistical approaches based framework (MSGAF) for wearable sensory gait data analysis. It has been applied to two clinical applications: discrimination of disturbed gait affected by neurodegenerative diseases and identification of patients with complex region pain syndrome (CRPS). The results demonstrate that: 1) It is feasible to discriminate gait patterns related three neurodegenerative diseases based on wearable sensory gait analysis. 2) It is feasible to assess the physical performance of patients with CRPS based on the analysis of accelerometry gait data. 3) Accelerometry gait data collected in a short distance can provide a large amount of information for gait monitoring in a home based environment. Two new feature selection algorithms have been proposed to find out the optimal feature set for a given condition related to gait disorder. A combination of gait features were proposed to analyse accelerometer based gait data. A device independent tool (i.e. iGAIT) has been developed to display and analyse gait acceleration data. It provides interactive functionality allowing users to manually adjust the analysis progress. A smartphone with an embedded accelerometer has been proposed to be a novel gait measurement technique. The utility and reliability of using a smartphone in gait pattern monitoring was studied. The impact of sampling frequency on the gait features extracted from accelerometer data was systematically investigated. The concept of contextual gait analysis was also proposed, which takes account of the impact of walking contexture and makes it possible to monitor gait pattern in a real life environment for a long term. Two data sets collected by an accelerometer were used, climbing stairs data and urban walking data.

This thesis looks at the problem of protecting large published statistical tables using cell suppression. Optimal cell suppression has been shown to be NP-Hard and can therefore only be applied to small tables. Using heuristic techniques to protect large tables tends to suppress far too many table cells lessening the utility of the table. Current state-of-the- art cell suppression algorithms can protect statistical tables with up to forty thousand cells. In this thesis a new model is derived that can fully protect statistical tables with up to one million cells without excessive over-suppression. This has been achieved by creating a new mathematical model that can protect cells in groups rather than individually. A pre-processing step was also introduced to reduce the number of cells that actually need to be protected. Further improvements have been gained through the employment of a self-adaptive Genetic Algorithm to optimise the order in which the groups of cells are protected and the employment of a surrogate fitness function to reduce execution time.