Chess is a complex logical game involving ongoing strategic forward planning and evaluation. Solving chess puzzles is one of the most common ways of training and developing chess skills. It involves continuing the game from a certain initial chessboard state against a real or AI opponent until defeat or a significant advantage is achieved. To ensure solving chess puzzles is efficient and engaging for the real player, it is important to understand the difficulty of the puzzle and match it with the skill of the solver. Accurate and fast assessment of puzzle difficulty is therefore a critical problem that online chess platforms need to solve at scale to optimally match many real players with adequate puzzles. To solve this challenge, we propose a chess knowledge-agnostic strategy to predict puzzle difficulty based solely on the moves of the players against their opponents trying to solve the puzzles. Specifically designed deep convolutional neural networks (CNN) were deployed as supervised learning predictors, fed with player moves represented as multichannel chessboard images. Extensive testing of our model with almost 4 million training examples against Glicko-2 evaluated puzzle difficulty ratings—considered as ground truth—resulted in good predictive performance. This was acknowledged by our runner-up result as the 7th place in the IEEE Big Data 2024 Cup and highlighted the capability of fast puzzle difficulty prediction based only on players’ moves as evidence, with no prior chess knowledge nor utilization of computationally expensive chess engines.

In the dynamic field of financial analytics, the ability to predict stock market trends is crucial for effective trading strategies, which is the task for FedCSIS 2024 Data Science Challenge: Predicting Stock Trends.This paper presents a comprehensive study on the use of hybrid gradient boosting models, incorporating both classification and regression approaches, to forecast stock trends across different sectors of the S&P 500.Utilizing a rich dataset comprising key financial indicators for 300 companies over a decade, our research aims to unravel the complexities of sector-specific trend predictions.The model leverages 58 financial indicators per company, along with their annual change metrics, to predict the future stock movements.In the preliminary phase of the competition, our hybrid model demonstrated promising results, achieving the lowest weighted error of 0.5941 among competitors.However, despite the initial success, the final phase of the model evaluation revealed a significant performance decline with the error rising above 0.84.This discrepancy highlights potential issues in model stability and preliminary performance when transitioning from a controlled to a truly unseen testing environment.This work not only underscores the complexities of predictive modeling in finance but also sets the stage for future research into creating more resilient AI-driven trading systems.

Food security relies on factors like availability, access, and stability, often assisted by food imports when local production falters. Importantly, these imports stabilize supplies, mitigate shortages and price volatility, and enhance economic stability. Anticipating import requirements is vital for proactive food security planning. In this case study, we employ multiple forecasting models to predict food import for a large number of products from multiple countries. The results highlight varying algorithm performance across datasets. Traditional statistical models remain highly competitive compared to newer alternatives, especially for shorter time series. Our study introduces a multi-model forecasting approach to predict periodic food imports, a pivotal tool for food authorities.

Muon scattering tomography (MST) allows to reconstruct internal structure of the matter based on imaging of muon rays, Coulomb scattering when passing near heavier nuclei of the penetrated matter. While point of closest approach (PoCA) and maximum-likelihood expectation maximization (MLEM) methods offer reasonable MST solutions, they require large number of muons or huge computational cost while still struggling to distinguish between similar materials. As an alternative we propose to utilize deep, fully convolutional neural networks with atrous convolutions designed to deliver robust semantic image segmentation (SIS). The information from muon rays is summarized as image pixels containing path length and momentum weighted statistics of per-voxel muon scattering angles and displacements encoded independently along image rgb channels. Reserving one channel for iterative self-improvement by feeding classification outputs back as inputs delivered significant material detection improvements with nearly perfect localization and improved material differentiation. The presented iterative SIS model has been tested within IEEE BigData 2023 Cup dedicated to muon scattering tomography and scored $2 ^{nd}$ place significantly outperforming PoCA/MLEM and Catboost with 99.9% localization accuracy and IoU score in excess of 0.35, on the very challenging task of detecting nested material layers composed of a mixture of up to 22 elements and compounds.

The rapid proliferation of Internet of Things (IoT) devices has revolutionized the way we interact with and manage our surroundings.However, this widespread adoption has also brought forth significant cybersecurity challenges.IoT devices, with their interconnectedness and varying functionalities, present a unique threat landscape that requires tailored detection techniques.Traditional approaches to cybersecurity, primarily focused on network monitoring and anomaly detection, often fall short in effectively identifying threats originating from IoT devices due to their dynamic and complex behaviors.This paper addresses our solution for FedCSIS 2023 Challenge: Cybersecurity Threat Detection in the behavior of IoT Devices.First, we aggregated time series features, and then at the feature selection stage, we filtered and combined different categorical and numerical features to generate four different feature sets.The Gradient boosting models, i.e. lightgbm, catboost and xgboost, are applied and trained individually with hyper-parameter tuning.The final three submissions are two best individual lightgbm models with the AUC scores of 0.9999 and 0.9998, respectively on the different feature sets, which secured the 4th place with a final score of 0.9993, and one ensemble result with a AUC score of 0.9998 from combination of xgboost, catboost and lightgbm, which has the final score of 0.9997 while unluckily was missing in the final three evaluation entries.

Gradient Boosting (GB) consistently outperforms other ML predictors especially in the context of binary classification based on multi-modal data of different forms and types.Its newest efficient implementations, including XGBoost, LGBM and CATBoost, push GB even further ahead with fast GPUaccelerated compute engine and optimized handling of categorical features.In an attempt to beat GB in both the performance and processing speed we propose a new simple yet fast and robust classification model based on predictive binning.At first all features undergo massively parallelized binning into a unified ordinally compressed risk representation, independently optimized to maximize the AUC score against the target.The resultant array of summarized micro-predictors, resembling 0-depth decision trees, directly expressing oridnally represented target risk, are then passed through the greedy feature selection to compose a robust wide-margin voting classifier, whose performance can beat GB while the extreme build and execution speed along with highly compressed representation welcomes extreme data sizes and realtime applicability.The model has been applied to detect cybersecurity attacks on IoT devices within FedCSIS'2023 Challenge and scored 2 nd place with the AU C ≈ 1, leaving behind all the latest GB variants in performance and speed.

A common business practice for transportation forwarders is to bid for shipping contracts at the transport or freight exchanges.Based on the detailed contract requirements they try to estimate the total expected cost of its execution and accordingly bid with the fixed price in advance for delivering such shipping service at the prescribed specification and schedule.The capability to accurately predict the cost of contract execution is the critical factor deciding about the profitability of offered shipping services as well as the amount of business drawn from freight exchanges.However, given highly volatile nature of the transport services ecosystem, it is difficult to simultaneously account for countless dynamically changing factors like fuel prices, currency exchange rates, temporal and spatial multitude of routing and implied traffic risks, the properties of cargo and shipping vehicles etc., which leads to big cost under-or overestimation resulting with loss-making contracts or equally painful missed revenue opportunities.In the context of FedCSIS 2022 data mining competition we propose an accurate and robust predictor of the cost of forwarding contracts built upon the detailed contract data using the ensemble of the state-of-theart gradient boosting-based regression models.Our established feature engineering framework combined with deep parametric optimization of the individual models and multi-faceted diversification techniques guiding hybrid final model ensembles were instrumental to outperform all the competitive predictors and win the FedCSIS 2022 contest.

Predicting the cost of forwarding contracts is a typical problem that logistics companies need to solve in order to optimize their business for a better profit.This is the challenge defined in the FedCSIS 2022 Competition where a fiveyear history of contract data and their delivery routes from a large Polish logistics company are provided to train a Machine Learning model.In addition to the contract data, historical wholesale fuel prices and euro exchange rates at the contract time are also provided.To address this challenge, we first designed a basic solution where we focused on feature engineering to find good impact features for the model.After that, the same set of features were used to train two different models: one using XGBoost and the other using LightGBM.The average predictions of the two boosting models were then used as the predictions for the next post-processing step.Finally, in the post-processing step, we designed and trained a simple linear regression model to capture the average monthly changes of the contract cost, given the changes of the fuel prices and euro exchange rates.These captured changes were used to post-process (adjust) the predictions in the previous step to address the issue that treebased models could not predict the value that they did not see before.While the basic solution with careful feature selection gave us a place in the top-5, our post-processing strategy in the last step helped us win the 3 rd prize in the competition.

When deploying a wireless sensor network (WSN) identifying optimal locations for the nodes is challenging, especially so if the environment is either (i) unknown or (ii) obstacle-rich. We proposes BISON (Bio-Inspired Self-Organizing Network), a variant of the Voronoi algorithm to address this challenge. Among the features of the approach are the restrictions placed upon the abilities of the nodes: (i) all information is sensed locally as well as (ii) subject to communication noise. Performance is measured as (i) the percentage of area covered, (ii) the total distance traveled by the nodes, (iii) the cumulative energy consumption and (iv) the uniformity of nodes’ distribution. Obstacle constellations and noise levels are studied systematically and a collision-free recovery strategy for failing nodes is proposed. Results obtained from extensive simulations show the algorithm outperforming previously reported approaches in both, convergence speed, as well as deployment cost.

An accurate evaluation of player-win likelihoods during online games is critical for controlling the attractiveness and immersion of the gameplay, especially if played against an AI bot. Predicting game victories is a very challenging problem heavily depends on the complexity and the stage of the game. With the multitude of ways and formats that the players, the gameplay, individual state and history can be encoded and utilized by machine learning (ML) models, it is fascinating to take part in the IEEE Big Data 2021 Cup and explore which data representations, feature engineering methods and prediction models work better than others and what levels of predictability they can achieve in the specific use case of predicting victories in Tactical Troops: Anthracite Shift online computer game. Our explorations throughout the game and a high 4th place in the Cup’s final indicate that a careful and comprehensive feature engineering methodology numerically capturing the last state of the game paired with the variants of the gradient boosting models offer a robust and competitive solution for this challenge.

At anytime in a tactics game, given the current status of the game players, it’s always interesting to predict who is going to win the game in the end. This is the challenge defined in the IEEE BigData Cup 2021, where the objective is to predict the likelihood of winning the game, Tactical Troops: Anthracite Shift, given information of the players as well as logged data of the game at the time the prediction needs to be made. To address this challenge, we have developed a hybrid model that combines a general model trained on all the data and two specific models trained on particular data of two distinct game modes. Our implementation was realized by LightGBM, one of the most popular tree-based libraries implementing the state-of-the-art Extreme Gradient Boosting algorithm. In addition to combined models, a thoroughly feature engineering and feature selection together with a special handling of missing values and categorical features, a careful hyper-parameter tuning, and a robust cross-validation strategy to avoid overfitting issues helped us win the 2ndplace in the competition.

Deep convolutional neural networks revolutionized the area of automated objects detection from images.Can the same be achieved in the domain of time series forecasting?Can one build a universal deep network that once trained on the past would be able to deliver accurate predictions reaching deep into the future for any even most diverse time series?This work is a first step in an attempt to address such a challenge in the context of a FEDCSIS'2020 Competition dedicated to network device workload prediction based on their historical time series data.We have developed and pre-trained a universal 3-layer bi-directional Long-Short-Term-Memory (LSTM) regression network that reported the most accurate hourly predictions of the weekly workload time series from the thousands of different network devices with diverse shape and seasonality profiles.We will also show how intuitive human-led post-processing of the raw LSTM predictions could easily destroy the generalization abilities of such prediction model.

The rapid expansion of Internet of Things (IoT) quickly evolves towards a connected network of AI-enabled, smart, multi-sensory devices that generate, consume and exchange enormous amounts of data. Stimulated by the omnipresent cloud-services and ever widening data bandwidth they aspire to control every aspect of our lives from transportation to our health, significantly increasing our reliance on the web-based services and their security. Already millions of cyber security alerts are generated every day and trigger increasingly costly investigations of the Security Operations Centres (SOC). In order to make their operation more efficient security warnings need to be reliably detected and classified based on various levels of severity, scale of potential damage or an ability to defend. We have responded to this challenge in the context of IEEE BigData Cup 2019 focused on predicting cyber security threats that require attention based on detailed logs of network activity leading to the security alert. We have developed a hybrid supervised learning ensemble model combining several state-of-the-art Extreme Gradient Boosting algorithms. Specifically, xGBoost and LightGBM model versions have been built on separate sets of features extracted from the raw logs of network events preceding the alerts and then synergically aggregated. Models' diversity imposed by algorithmic differences, complementary feature subsets, and individually optimized hyperparameters, combined with robust stratified cross-validation scheme resulted with the best true alerts detection rate yielding the AUC score in excess of 0.93, that outperformed all other 248 competing teams.

Support Vector Regression (SVR) as a supervised machine learning algorithm have gained popularity in various fields.However, the quadratic complexity of the SVR in the number of training examples prevents it from many practical applications with large training datasets.This paper aims to explore efficient ways that maximize prediction accuracy of the SVR at the minimum number of training examples.For this purpose, a clustered greedy strategy and a Genetic Algorithm (GA) based approach are proposed for optimal subset selection.The performance of the developed methods has been illustrated in the context of Clash Royale Challenge 2019, concerned with decks' win rate prediction.The training dataset with 100,000 examples were reduced to hundreds, which were fed to SVR training to maximize model prediction performance measured in validation R 2 score.Our approach achieved the second highest score among over hundred participating teams in this challenge.

Support Vector Regression (SVR) is a powerful supervised machine learning model especially well suited to the normalized or binarized data.However, its quadratic complexity in the number of training examples eliminates it from training on large datasets, especially high dimensional with frequent retraining requirement.We propose a simple two-stage greedy selection of training data for SVR to maximize its validation set accuracy at the minimum number of training examples and illustrate the performance of such strategy in the context of Clash Royale Challenge 2019, concerned with efficient decks' win rate prediction.Hundreds of thousands of labelled data examples were reduced to hundreds, optimized SVR was trained on to maximize the validation R 2 score.The proposed model scored the first place in the Cash Royale 2019 challenge, outperforming over hundred of competitive teams from around the world.

Designing a robust and adaptable Artificial Intelligence (AI) opponent in a computer game would ensure the game continues to challenge, immerse and excite the players at any stage.The outcomes of card based games such as "Heartstone: Heroes of Warcraft", aside the player skills, heavily depend on the initial composition of player card decks.To evaluate this impact we have developed a new robust regression network in a context of the AAIA Data Mining Competition 2018, which tries to predict the average win-rates of the specific combinations of bot-player and card decks.Our network is composed of 2 levels: the entry level with an array of finely optimized state of the art regression models including Extreme Learning Machines (ELM), Extreme Gradient Boosted decision tree (XGBOOST), and Least Absolute Shrinkage and Selection Operator (LASSO) regression trained via supervised learning on the labeled training dataset; and just a single ELM at the 2 nd level installed to learn to correct the predictions from the 1 st level.The final solution received the root of the mean squared error (RMSE) of just 5.65% and scored the 2 nd place in AAIA'2018 competition.This paper also presents two other runner-up models receiving RMSE of 5.7% and 5.86%, scoring the 4 th and the 6 th place respectively.

Success of many computer games depends on designing a robust and adaptable AI opponent that would ensure the games continue to challenge, immerse and excite the players at any stage.The outcomes of card based games like "Heartstone: Heros of Warcraft", aside the player skills heavily depend on the initial composition of player card decks.To evaluate this impact we have developed an ensemble prediction model that tries to predict the average win-rates of the specific combination of bot-player and card decks.Our ensemble model consists of three sub-models: two Logistic Regression models and one Deep Learning model.The models are trained with both provided data and additional data about the cards, their health, attack power and cost.To avoid overfitting, we employ a trick to generate predictions for all possible combinations of opponent players and decks and obtain the result as the average of all these predictions.

This submission represents a proposition of the Full Paper in the Research-to-Practice category in which we present our innovative work on automated student knowledge mapping and gaps discovery.Students' ability to learn and retain new knowledge depends on the precise state of their current knowledge distributed along a complex network of related concepts. Uneven distribution of competencies in the required knowledge concepts, or even worse - knowledge gaps - in certain critical concepts, may significantly hinder students' ability to acquire knowledge efficiently and prevent them from maximizing the potential of their talents. Discovery of and bridging these gaps at the individual student and group level is critically important for students to maximize their learning efficiency, for the educators to maximize the impact of the teaching, and for the governments to maximize the effectiveness of their educational programs and policies.We propose a novel methodology for automated student knowledge mapping and gaps discovery from the results of multiple choice test questions extracted from a single .pdf report through pattern recognition techniques. Automatically lifted test performance data are then decomposed into student competencies in the related concepts required to correctly answer the test questions. Simple constrained linear optimization is deployed to find the optimal student competencies in the related concepts such that their linear combination best reconstructs the achieved test scores. Extracted knowledge maps in the related concepts allowed to discover common concept gaps for individual students and their groups, which has been successfully illustrated in the context of mathematics preparatory course at the university level with hundreds of students and thousands of test questions.

In the area of applied optimisation, heuristics are a popular means to address computational problems of high complexity.Modelling the problem and mapping all variations of its solution into a so-called solution space are integral parts of this process.Representing solutions as graphs is common and, for a special type of graph, Pr üfer Code (PC) offers a computationally efficient mapping (algorithms of Θ(n)-complexity are known) to n-2 dimensional Euclidean space.However, this encoding does not preserve properties such as e.g.locality and therefore PC has been shown to be a bad choice for entire classes of problems.We argue that PC does allow the preservation of some properties (e.g.degree of branching and branching vertices) and that these are sufficiently relevant for certain types of problems to motivate encoding them in PC.We present our investigations and provide an example where PC has been shown to be a useful encoding.

This paper introduces an ensemble model that solves the binary classification problem by incorporating the basic Logistic Regression with the two recent advanced paradigms: extreme gradient boosted decision trees (xgboost) and deep learning.To obtain the best result when integrating sub-models, we introduce a solution to split and select sets of features for the sub-model training.In addition to the ensemble model, we propose a flexible robust and highly scalable new scheme for building a composite classifier that tries to simultaneously implement multiple layers of model decomposition and outputs aggregation to maximally reduce both bias and variance (spread) components of classification errors.We demonstrate the power of our ensemble model to solve the problem of predicting the outcome of Hearthstone, a turn-based computer game, based on game state information.Excellent predictive performance of our model has been acknowledged by the second place scored in the final ranking among 188 competing teams.

Global demographic trends place an increasing demand on the efficient and customizable design of large scale buildings and digital services offered within. Indoor wireless access has become standard and ensuring a thorough and economically sound wireless signal coverage throughout the building is not a trivial problem. An In-building Distributed Antenna System (IB-DAS) extends the wireless access from the base station to distributed antennas through a complex network of coaxial cables and power splitters. For high rise buildings and other multi-unit complexes, the initial cost of IB-DAS (cabling and splitters) and the running costs of powering the network are quite significant, hence the optimal design of an IB-DAS network is critical. Our proposition here is a novel In-Building DAS optimization model that utilizes Particle Swarm Optimization (PSO) to provide near optimal network topology that simultaneously minimizes the cost of the cabling and equipment for the whole building. Our PSO model uses Prufer number representation to efficiently traverse through different spanning tree solutions. As opposed to other approaches which solve this problem only partially (either equipment or cabling), typically considering a building as a single floor unit, our approach is a complete proposition for DAS optimization that is scalable and robust to deliver IB-DAS designs even for the tallest buildings beyond one hundred floors. We demonstrate that our model is capable of obtaining optimal solutions for small buildings and near optimal solutions for tall buildings.

Social media have become an important source of data and can provide near-instantaneous information which can be analysed to generate predictive models and to support decision making. Much work has been done in short message analysis such as trend analysis, short message classification, etc. However, to generate an accurate and concise conclusion/assertion from all the relevant information remains challenging. In this paper we propose a method to analyse microblog messages at both ‘word/term’ level and ‘concept’ level to generate assertions accurately and instantly. To analyse the concept level, we define a small seed ontology which is a semi-automatically generated extension of an existing ontology. By doing this we achieve both accurate assertions and avoid the costly overhead of defining the whole knowledgebase manually. We then use the proposed method to make traffic assertions from a microblog stream to demonstrate the advantages of the approach.

Sequential, predominantly temporal nature of the vast amounts of big data released every day from many different sources could potentially be linked, aligned along the time and deliver new evidence for the next generation predictive systems or knowledge discovery engines. However, big data owners are reluctant to share their data due to legally binding privacy and identity protection concerns, thereby posing a major hurdle preventing shared exploitation of big data on a massive scale. Data anonymization is expected to solve this problem, yet the current approaches are limited predominantly to univariate time series generalized by aggregation or clustering to eliminate identifiable uniqueness of individual data points or patterns. For multivariate time series, uniqueness among of the combination of values or patterns across multiple dimensions is much harder to eliminate due the to exponentially growing number of unique configurations of point values across multiple dimensions. Our method implements linearly scalable asynchronous summarization of multivariate time series independently at every dimension. As a result the series retain only a small subset of defining points at different times along multiple dimensions effectively breaking up the multivariate time series into a collection of summarized univariate time series that are perturbed from the original series in terms of actual points and pattern shapes. Current implementation of the anonymizing summarization involves shape preserving greedy elimination and aggregation that supports parallel cluster processing for big data implementation.

Collaborative Learning (CL) has been considered as an effective way to improve the learning outcomes of students in contrast to individual learning. However, assigning a groupwork task to a team of students does not guarantee a successful performance, and in fact could hinder the benefits of group learning if the members do not interact as expected. Indeed, group learning performance is largely dependent on group composition. In this work we address the challenge of identifying the characteristics of the individual group members that bare the significant impact on the performance of the groupwork. Specifically we investigate the impact that a combination of individual student performances and gender have on the group performance and intend to find generic segmentation guidelines that would map smoothly onto the groupwork performance. A novel grouping method is proposed, which splits the set of students into groups that maximize one of the two desired criteria: the expected average groupwork performance or the average improvement achieved by a student as a result of synergic group learning and interaction effects. The model uses global optimization approach to identify optimal students allocation into the groups that best satisfy the optimization criteria. We illustrate our findings on the data obtained from the trial of the Collaborative Learning Environment (CLE) software. The CLE was developed at Etisalat British Telecom Innovation Centre (EBTIC) and tested over one semester with a sample of 122 students working in different groups in the Engineering and Molecular Biology courses at Khalifa University. The results of our method can not only help to understand the significant factors impacting group performance in group-based learning, but can also provide practical strategies on optimal group composition for collaborative learning activities.

The explosive growth of big data poses a processing challenge for predictive systems in terms of both data size and its dimensionality.Generating features from text often leads to many thousands of sparse features rarely taking non-zero values.In this work we propose a very fast and robust feature selection method that is optimised with the Naive Bayes classifier.The method takes advantage of the sparse feature representation and uses diversified backward-forward greedy search to arrive with the highly competitive solution at the minimum processing time.It promotes the paradigm of shifting the complexity of predictive systems away from the model algorithm, but towards careful data preprocessing and filtering that allows to accomplish predictive big data tasks on a single processor despite billions of data examples nominally exposed for processing.This method was applied to the AAIA Data Mining Competition 2014 concerned with predicting human injuries as a result of fire incidents based on nearly 12000 risk factors extracted from thousands of fire incident reports and scored the second place with the predictive accuracy of 96%.

The global educational landscape is changing due to the advancement of technologies and services, some have termed it as the "climate change" in education. The students of today engage with the learning environment differently from the students of yesterday. The traditional landscape is often perceived as "formal", "passive", "direct", and "push" learning environment designed largely for the knowledge consumers, and the modern landscape is often perceived as "informal", "active", "collaborative", "social", and "pull" learning environment designed not only for the knowledge consumers but also for the knowledge creators. This paradigm shift in education is imminent and has gathered a lot of interests in the recent years to create and/or adapt the education environment for the 21st century. To facilitate this impending modern landscape, this paper would introduce a new radical construct, termed as self-organized and/or purposely-organized peer-to-peer learning, and its impacts towards both the formal and informal learning environment. An overview of some of the solutions, that have been developed under the holistic intelligent campus (iCampus) initiative, would also be delineated so as to address such aspectual issues and challenges pertaining to the next generation smart learning environment.

Rapidly evolving businesses generate massive amounts of time-stamped data sequences and cause a demand for both univariate and multivariate time series forecasting. For such data, traditional predictive models based on autoregression are often not sufficient to capture complex nonlinear relationships between multidimensional features and the time series outputs. In order to exploit these relationships for improved time series forecasting while also better dealing with a wider variety of prediction scenarios, a forecasting system requires a flexible and generic architecture to accommodate and tune various individual predictors as well as combination methods. In reply to this challenge, an architecture for combined, multilevel time series prediction is proposed, which is suitable for many different universal regressors and combination methods. The key strength of this architecture is its ability to build a diversified ensemble of individual predictors that form an input to a multilevel selection and fusion process before the final optimized output is obtained. Excellent generalization ability is achieved due to the highly boosted complementarity of individual models further enforced through cross-validation-linked training on exclusive data subsets and ensemble output postprocessing. In a sample configuration with basic neural network predictors and a mean combiner, the proposed system has been evaluated in different scenarios and showed a clear prediction performance gain.

Automated pedestrian detection is a forward looking challenge for future driver support systems in automotive industry. Such system would have to make safety critical decisions based on poor quality images shot in real-time from the unstable moving vehicles. The proposed system offers a simple yet very effective detection methodology based on mixture of Gaussians (MoG) aided by an Expectation-Maximisation (EM) clustering algorithm. The algorithm operates on a number of features built by aggregation of different variations of the first and second order pixel gradients related to the aggregated templates of pedestrian and non-pedestrian classes. For each class the algorithm fits a fixed number of clusters and using Gaussian kernels optimises the parameters of the Gaussian Mixture model such that the probabilities of belonging to the intraclass clusters is maximised. Given a new image the system instantly generates relative features and uses mixture model to build posterior probability densities for all clusters and after aggregation and renormalisation, posterior class probabilities. The system has been fine-tuned against its parameters and feature subsets and tested using almost 10000 real images provided by DaimlerChrysler. Reaching the testing performance in excess of 95% the model was announced the winner of the NISIS Competition 2007.

Rapidly evolving businesses generate massive amounts of time-stamped data sequences and defy a demand for massively multivariate time series analysis. For such data the predictive engine shifts from the historical auto-regression to modelling complex non-linear relationships between multidimensional features and the time series outputs. In order to exploit these time-disparate relationships for the improved time series forecasting, the system requires a flexible methodology of combining multiple prediction models applied to multiple versions of the temporal data under significant noise component and variable temporal depth of predictions. In reply to this challenge a composite time series prediction model is proposed which combines the strength of multiple neural network (NN) regressors applied to the temporally varied feature subsets and the postprocessing smoothing of outputs developed to further reduce noise. The key strength of the model is its excellent adaptability and generalisation ability achieved through a highly diversified set of complementary NN models. The model has been evaluated within NISIS Competition 2006 and NN3 Competition 2007 concerning prediction of univariate and multivariate time-series. It showed the best predictive performance among 12 competitive models in the NISIS 2006 and is under evaluation within NN3 2007 Competition.