.

With the constant increase in smartphone sales, integrated sensors have becomeavailable to the average user. This allows for mobile applications to utilise theuser’s context to provide more accurate information. The popularity of smartphones also attract developers to create audio functionalities that have earlier been restricted to calling interfaces. There is an increasing interest for Automatic Speech Recognition (ASR) services aimed at everyday tasks, where Apple’s release of SIRI is a good example of a system that has contributed to the gained popularity. This report describes TaleTUC, a proof of concept system for the domain of bus route information. TaleTUC uses ASR combined with context-awareness through Case-based Reasoning (CBR), to recognise spoken bus stop names. It is built on a client-server architecture, where theTABuss (Marcussen and Eliassen, 2011) Android application has been extendedto operate as a client. As a TaleTUC client, TABuss uses speech as input to itsmain query functionality, which provides bus route suggestions through BusTUC and AtB’s real-time system. Three modules have been developed server-side, where one is used for ASR, and the two others are used for context-awareness. Testing of the three modulescombined showed improved results compared to the ASR module alone, which indicates that context-awareness is a suitable technology to combine with ASR.

In this report we introduce a classification system named Grouping of Features (GoF), together with a theoretical exploration of some of the important concepts in the Instant Based Learning(IBL)-field that are related to this system.A dataset's original features are by the GoF-system grouped together into abstract features. Each of these groups may capture inherent structures in one of the classes in the data. A genetic algorithm is used to extract a tree of such groups that can be used for measuring similarity between samples. As each class may have different inherent structures, different trees of groups are found for the different classes. To adjust the importance of one group in regards to the classifier, the concept of power average is used. A group's power-average may let either the smallest or the largest value of its group dominate, or take any value in-between. Tests show that the GoF-system outperforms kNN at many classification tasks.The system started as a research project by Verdande Technology, and a set of algorithms had been fully or partially implemented before the start of this thesis project. There existed no documentation however, so we have built an understanding of the fields on which the system relies, analyzed their properties, documented this understanding in explicit method descriptions, and tested, modified and extended the original system.During this project we found that scaling or weighting features as a data pre-processing step or during classification often is crucial for the performance of the classification-algorithm. Our hypothesis then was that by letting the weights vary between features and between groups of features, more complex structures could be captured. This would also make the classifier less dependent on how the features are originally scaled. We therefore implemented the Weighted Grouping of Features, an extension of the GoF-system.Notable results in this thesis include a 95.48 percent and 100.00 percent correctly classified non-scaled UCI Wine dataset using the GoF- and WGoF-system, respectively.

AbstractThe master thesis focuses on ensemble approaches applied to intrusion detection systems (IDSs). The ensemble approach is a relatively new trend in artificial intelligence in which several machine learning algorithms are combined. The main idea is to exploit the strengths of each algorithm of the ensemble to obtain a robust classifier. Moreover, ensembles are particularly useful when a problem can be segmented into subproblems. In this case, each module of the ensemble, which can include one or more algorithms, is assigned to one particular subproblem. Network attacks can be divided into four classes: denial of service, user to root, remote to local and probe. One module of the ensemble designed in this work is itself an ensemble of decision trees and is specialized on the detection of one class of attacks. The inner structure of each module uses bagging techniques to increase the accuracy of the IDS. Experiments showed that IDSs obtain better results when each class of attacks is treated as a separate problem and handled by specialized algorithms. This work have also concluded that these algorithms need to be trained with specific subsets of fea- tures selected according to their relevance to the class of attack being detected. The efficiency of ensemble approaches is also highlighted. In all experiments, the ensemble was able to bring down the number of false positives and false negatives. However, we also observed the limitations of the KDD99 dataset. In particular, the distribution of examples of remote to local attacks between the training set and test set made difficult the evaluation of the ensemble for this class of attack.

Careful architectural decisions are required in order to create a highly available and scalable search system. This requires an in-depth analysis and understanding of the architecture and context of each deployment. Different requirements placed upon the system by different deployments mean different solutions provide the best case by case result, thus benchmarks provide an invaluable source of information.This thesis provides an overview of common components and important aspects of a distributed search system. It then gives an overview of different partitioning techniques before going into the details of repartitioning and rebalancing in a document-partitioned full-text search system.A processing framework that draws inspiration from flow-based programming literature is introduced, which is shown a valuable tool in creating custom tailored search solutions. The implementation is used to benchmark different repartitioning and rebalancing strategies.In conclusion, the techniques mentioned in the thesis show great promise in creating custom, maintainable and flexible partitions. The processing framework enables each specific deployment to easily compare different partitioning schemes and associated manageability and maintenance costs to determine the best fit for any given situation.

Texas Hold'em poker provides an interesting test-bed for AI research with characteristics such as uncertainty and imperfect information, which can also be found in domains like medical decision making. Poker introduces these characteristics through its stochastic nature and limited information about other players strategy and hidden cards. This thesis presents the development of a Bayesian Case-based Reasoner for Poker (BayCaRP). BayCaRP uses a Bayesian network to model opponent behaviour and infer information about their most likely cards. The case-based reasoner uses this information to make an informed betting decision. Our results suggests that the two reasoning methodologies combined achieve a better performance than either could on its own.

This thesis investigates the prediction of possible stock price changes immediately after news article publications, by automatic analysis of these news articles. Some background information about financial trading theory and text mining is given in addition to an overview of earlier related research in the field of automatic analyzes of news articles for predicting future stock prices. In this thesis a system is designed and implemented to predict stock price trends for the time immediately after the publication of news articles. This system consists mainly of four components. The first component gathers news articles and stock prices automatically from internet. The second component prepares the news articles by sending them to some document preprocessing steps and finding relevant features before they are sent to a document representation process. The third component categorizes the news articles into predefined categories, and finally the fourth component applies appropriate trading strategies depending on the category of the news article. This system requires a labeled data set to train the categorization component. This data set is labeled automatically on the basis of the price trends directly after the news article publication. An additional label refining step using clustering is added in an attempt to improve the labels given by the basic method of labeling by price trends.The findings indicate that a categorization of news articles provides additional information that can be used to forecast stock price trends. Experiments showed that the label refining method greatly improves the performance of the system. It was also shown that the timing of when to start the price trends used to label the data sets had a significant impact on the results. Trading simulations performed with the systems managed to gain positive returns (profits) on most of its trades. Some of the methods also managed to give better results than what trades performed with the manually labeled data set did.

This thesis presents a graph-based approach to the problem of text representation. The work is motivated by the need for better representations for use in textual Case-Based Reasoning (CBR). In CBR new problems are solved by reasoning based on similar past problem cases. When the cases are represented in free text format, measuring the similarity between a new problem and previously solved problems become a challenging task. The case documents need to be re-represented before they can be compared/matched.Textual CBR (TCBR) addresses this issue. We investigate automatic re-representation of textual cases, in particular measuring the salience of features (entities in the text) towards this end. We use the classical vector space model in Information Retrieval (IR) but investigate whether graph-representation and salience inference using graphs can improve on the Term Frequency (TF) and Term Frequency-Inverse Document Frequency (TF-IDF) measures, emph{bag of words} approaches predominant in IR.Our special focus is whether, and possibly how, the co-occurrence and the syntactic dependency relations between terms have an impact on feature weighting. We measure salience through the notion of graph centrality. We experiment with two types of application tasks, classification and case retrieval. Although classification is not a typical TCBR task, it is easier to find datasets for this application, and the centrality measures we have studied are not specific to TCBR. The experiments on this task are therefore relevant to the second application task which is our ultimate target. We test various centrality metrics described in the literature, make a distinction between local and global weighting measures and compare them for both application tasks. In general, our graph-based salience inference methods perform better than TF and TF-IDF.

Electroencephalography (EEG) equipment are becoming more available on thepublic market, which enables more diverse research in a currently narrow field.The Brain-Computer Interface (BCI) community recognize the need for systemsthat makes BCI more user-friendly, real-time, manageable and suited for peoplethat are not forced to use them, like clinical patients, and those who are disabled.Thus, this project is an effort to seek such improvements, having a newly availablemarket product to experiment with: a single channel brain wave reader. However,it is important to stress that this shift in BCI, from patients to healthy and ordinaryusers, should ultimately be beneficial for those who really need it, indeed.The main focus have been building a system which enables usage of the availableEEG device, and making a prototype that incorporates all parts of a functioningBCI system. These parts are 1) acquiring the EEG signal 2) process and classify theEEG signal and 3) use the signal classification to control a feature in a game. Thesolution method in the project uses the NeuroSky mindset for part 1, the Fouriertransform and an Artificial Neural Network for classifying brain wave patterns inpart 2, and a game of Snake uses the classification results to control the characterin part 3.This report outlines the step-by-step implementation and testing for this system,and the result is a functional prototype that can use user EEG to control the snakein the game with over 90% accuracy. Two mental tasks have been used to separatebetween turning the snake left or right, baseline (thinking nothing in particular)and mental counting. The solution differentiates from other appliances of the NeuroSkymindset that it does not require any pre-training for the user, and it is onlypartially real-time.

Pythia is a programming game that allows the player to change pieces of theirenvironment through use of the programming language Python. The idea is that thegame could be used as a part of teaching simple programming to first year universitystudents. The game should be fun enough for the students to keep playing, teachenough for it to earn a place as a teaching tool, and it should be usable by allstudents. It should also be possible for a teacher to create their own content for thegame.Pythia was implemented by extending the Python-interpreter Jython and building a game around it. The game was rendered using a simple hardware accelerationlibrary. A simple story was invented and there was some research on learning andprogramming in games.A set of levels was made, matching the story and introducing puzzles related tosimple programming. These levels were used in testing to collect data on usability,entertainment, and learning. There were also tests of the performance of the gameon several systems, and an evaluation was made on creating content for the game.The game has potential for being used to teach programming to first yearstudents, as testers found it to be both fun and educational. We do not know if itwould be possible to use it, as it does not currently run on thin clients. If studentscan run it, we feel that it should be possible for teachers to create puzzles thatemulate the teaching goal.