Mohamed Abou-Zleikha

Evolving Random Forest for Preference Learning

This paper introduces a novel approach for pairwise preference learning through a combination of an evolutionary method and random forest. Grammatical evolution is used to describe the structure of the trees in the Random Forest (RF) and to handle the process of evolution. Evolved random forests are evaluated based on their efficiency in predicting reported preferences. The combination of these two efficient methods for evolution and modelling yields a powerful technique for learning pairwise preferences. To test the proposed methodology and compare it to other methods in the literature, a dataset of 1560 sessions with detail information about user behaviour and their self-reported preferences while interacting with a game is used for training and evaluation. The method demonstrates ability to construct accurate models of user experience from preferences, behavioural and context data. The results obtained for predicting pairwise self-reports of users for the three emotional states engagement, frustration and challenge show very promising results that are comparable and in some cases superior to those obtained from state-of-the-art methods.

A Progressive Approach to Content Generation

PCG approaches are commonly categorised as constructive, generate-and-test or search-based. Each of these approaches has itsdistinctive advantages and drawbacks. In this paper, we propose an approach to Content Generation (CG) – in particular level generation – that combines the advantages of constructive and search-based approaches thus providing a fast, flexible and reliable way of generating diverse content of high quality. In our framework, CG is seen from a new perspective which differentiates between two main aspects of the gameplay experience, namely the order of the in-game interactions and the associated level design. The framework first generates timelines following the search-based paradigm. Timelines are game-independent and they reflect the rhythmic feel of the levels. A progressive, constructive-based approach is then implemented to evaluate timelines by mapping them into level designs. The framework is applied for the generation of puzzles for the Cut the Rope game and the results in terms of performance, expressivity and controllability are characterised and discussed.

Audio-Visual Classification of Sports Types

In this work we propose a method for classification of sports types from combined audio and visual features extracted from thermal video. From audio Mel Frequency Cepstral Coefficients (MFCC) are extracted, and PCA are applied to reduce the feature space to 10 dimensions. From the visual modality short trajectories are constructed to represent the motion of players. From these, four motion features are extracted and combined directly with audio features for classification. A k-nearest neighbour classifier is applied for classification of 180 1-minute video sequences from three sports types. Using 10-fold cross validation a correct classification rate of 96.11% is obtained with multimodal features, compared to 86.67% and 90.00% using only visual or audio features, respectively.

A Projection-Based Approach for Real-Time Assessment and Playability Check for Physics-Based Games

This paper introduces an authoring tool for physics-based puzzle games that supports game designers through providing visual feedback about the space of interactions. The underlying algorithm accounts for the type and physical properties of the different game components. An area of influence, which identifies the possible space of interaction, is identified for each component. The influence areas of all components in a given design are then merged considering the components’ type and the context information. The tool can be used offline where complete designs are analyzed and the final interactive space is projected, and online where edits in the interactive space are projected on the canvas in realtime permitting continuous assistance for game designers and providing informative feedback about playability.

Automatic Analysis of Activities in Sports Arenas Using Thermal Cameras

The demand for automatically gathered data is a societal trend quickly extending to all aspects of human life. Knowledge on the utilization of public facilities is of interest for optimising use and cutting expenses for the owners. Manual observations are both cumbersome and expensive, and they have a risk of incorrect results due to subjective opinions or lack of interest in the given task. In this paper we present the main results of a 5-year long research project revolving around the real-world application of automatic analysis of activities in sports arenas. Three topics are explored: Counting people, recognising activities, and estimating energy expenditure. The project is based on thermal image data, to preserve privacy while capturing video in public sports arenas. This paper aim to provide an overview of our published methods and results within these three topics and add a discussion of the results and perspectives of this research.

Projecting emotional speech into arousal-valence space using pairwise preference learning

Emotion recognition in speech is a very challenging task in the speech processing domain. Because of the continuity characteristics of human emotion, most of the recent research focuses on recognising emotion in a continuous space. While previous attempts for speech emotion annotation adopted the likert-like scaling system in a continuous space and relied on prediction models to predict emotion we, in this research, propose a new method for data labelling based on a pairwise data annotation. A set of constraints was proposed to decrease the number of pairs required to label. The annotated data is used to construct a regression model using the pairwise evolutionary multivariate adaptive regression spline method. The experiments performed show high recognition accuracies compared to the baseline random pairwise assignment.

Non-linguistic vocal event detection using online random forest

Accurate detection of non-linguistic vocal events in social signals can have a great impact on the applicability of speech enabled interactive systems. In this paper, we investigate the use of random forest for vocal event detection. Random forest technique has been successfully employed in many areas such as object detection, face recognition, and audio event detection. This paper proposes to use online random forest technique for detecting laughter and filler and for analyzing the importance of various features for non-linguistic vocal event classification through permutation. The results show that according to the Area Under Curve measure the online random forest achieved 88.1% compared to 82.9% obtained by the baseline support vector machines for laughter classification and 86.8% to 83.6% for filler classification.

Utilising Tree-Based Ensemble Learning for Speaker Segmentation

In audio and speech processing, accurate detection of the changing points between multiple speakers in speech segments is an important stage for several applications such as speaker identification and tracking. Bayesian Infor- mation Criteria (BIC)-based approaches are the most traditionally used ones as they proved to be very effective for such task. The main criticism levelled against BIC-based approaches is the use of a penalty parameter in the BIC function. The use of this parameters consequently means that a fine tuning is required for each variation of the acoustic conditions. When tuned for a certain condition, the model becomes biased to the data used for training limiting the models general- isation ability. In this paper, we propose a BIC-based tuning-free approach for speaker seg- mentation through the use of ensemble-based learning. A forest of segmentation trees is constructed in which each tree is trained using a sampled version of the speech segment. During the tree construction process, a set of randomly selected points in the input sequence is examined as potential segmentation points. The point that yields the highest ΔBIC is chosen and the same process is repeated for the resultant left and right segments. The tree is constructed where each node corresponds to the highest ΔBIC with the associated point index. After building the forest and using all trees, the accumulated ΔBIC for each point is calcu- lated and the positions of the local maximums are considered as speaker changing points. The proposed approach is tested on artificially created conversations from the TIMIT database. The approach proposed show very accurate results compara- ble to those achieved by the-state-of-the-art methods with a 9% (absolute) higher F1 compared with the standard ΔBIC with optimally tuned penalty parameter.