In this thesis we will construct a recommender system for course selection in higher education (more specifically, at The Norwegian University of Science and Technology). Some of what makes our approach novel compared with existing solutions is that we weight each user in the collaborative filtering process based on their chosen degree subject (major) and wether or not the two users being compared are friends. Also we utilize both collaborative filtering and content-based recommendations in a hybrid solution. Another novel aspect of our solution is that we construct our system on top of an existing website for rating courses. This gives us unique access to a dataset containing thousands of user-ratings of courses.

In the field of artificial intelligence, recommender systems are methods for predicting the relevance items to a users. The items can be just about anything, for example documents, articles, movies, music, events or other users. Recommender systems examine data such as ratings, query logs, user behavior and social connections to predict what each user will think of each item.Modern recommender systems combine multiple standard recommenders in order to leverage disjoint patterns in available data. By combining different methods, complex predictions that rely on much evidence can be made. These aggregations can for example be done by estimating weights that result in an optimal combination.However, we posit these systems have an important weakness. There exists an underlying, misplaced subjectivity to relevance prediction. Each chosen recommender system reflects one view of how users and items should be modeled. We believe the selection of recommender methods should be automatically chosen based on their predicted accuracy for each user and item. After all, a system that insists on being adaptive in one particular way is not really adaptive at all.This thesis presents a novel method for prediction aggregation that we call adaptive recommenders. Multiple recommender systems are combined on a per-user and per-item basis by estimating their individual accuracy in the current context. This is done by creating a secondary set of error estimating recommenders. The core insight is that standard recommenders can be used to estimate the accuracy of other recommenders. As far as we know, this type of adaptive prediction aggregation has not been done before.Prediction aggregation (combining scores) is tested in a recommendation scenario. Rank aggregation (sorting results lists) is tested in a personalized search scenario. Our initial results are promising and show that adaptive recommenders can outperform both standard recommenders and simple aggregation methods. We will also discuss the implications and limitations of our results.

The goal of this thesis was to investigate the usage of mobile devices with multitouch capabilities in the learning of procedural knowledge. A system, consisting of three prototypes, was to be implemented as a way of examining our two hypotheses:H1: Through using a conceptual model close to how the human mind perceive objects, we can increase consistency both in the creation of new user manuals and in the learning process.H2: By taking advantage of multitouch technologies we can introduce a more natural way of interacting on virtual representations of real-life objects.A lot of research was conducted on the usage of a conceptual model containing information on the physical attributes and the procedural knowledge to back our applications, and how this best could be realized. Existing technologies for creating 3D models was investigated, but was quickly discarded due to the unique representation that was needed to successfully integrate the model with GOMS. The research process concluded that an application for describing new devices would have to be developed as well.Three applications was developed to investigate our hypotheses: an application for describing the aspects of a device, written for Mac OS, a server for communicating with prolog over TCP, written in Java, and an application for displaying the device and allowing for interaction, written for the iOS platform. The final versions of these three prototypes made it possible to create objects consisting of cubes, storing them on the server, and rendering them on the mobile application. The report concludes by discussing the utility of our prototype in regards to the hypotheses. Although not in its optimal state, the prototype demonstrates the utility of pure gestural interfaces, and how well established technologies such as prolog and GOMS can be used to empower them. Finally, interesting extensions and further work based on this thesis is proposed, demonstrating its versatility.

Artificial Evolution has shown great potential in the musical domain. One task in which Evolutionary techniques have shown special promise is in the automatic creation or composition of music. However, a major challenge faced when constructing evolutionary music composition systems is finding a suitable fitness function.Several approaches to fitness have been tried. The most common is interactive evaluation. However, major efficiency challenges with such an approach have inspired the search for <i>automatic</i> alternatives.In this thesis, a music composition system is presented for the evolution of novel melodies. Motivated by the repetitive nature of music, a <i>quantitative</i> approach to automatic fitness is pursued. Two techniques are explored that both operate on frequency distributions of musical events. The first builds on <i>Zipf's Law</i>, which captures the scaling properties of music. Statistical <i>similarity</i> governs the second fitness function and incorporates additional domain knowledge learned from existing music pieces.Promising results show that pleasant melodies can emerge through the application of these techniques. The melodies are found to exhibit several favourable musical properties, including rhythm, melodic locality and motifs.

In this thesis, I present an approach to a case-based reasoning system with explanation capabilities in the Warhammer Fantasy Battle domain. This product is meant to support Warhammer gamers in their initial army lineup, by providing suggestions based on previously successful games against an opposing horde. Explanations will be used in order to convey the reasoning behind the solution, to present the data the solution is based upon and why certain changes were made.The created product is capable of creating the army lineup and give partially satisfactory explanations, based on the goals set both for the application as a whole and explanations. Although a full domain model is not implemented, are the results promising; with the inclusion of more domain knowledge and cases, will a fully competent and accurate system be achievable.

When tasks traditionally performed by humans are automated it is important thatthe machines are able to communicate how these tasks are solved and why. Whena user is surprised by the point of time where the task is executed, there is a needto be able to get an explanation to why this point in time was chosen.This project aims at investigating how intelligent systems in general, and case-based reasoning systems in particular can become explanation-aware. Our aim isprimarily to investigate existing case-based reasoning systems to see if explanation-awareness is achievable. Secondary, our aim is to develop a simple case-based rea-soning engine that complies with our theoretical work on explanation-awareness.

Artificial intelligence and learning is a growing field. There are many ways of making a computer program learn, in most cases one have a specific problem one wants to solve and do not really care how it is solved. This thesis have a specific problem, but the main focus is on how it is solved. One of the most exciting ways to learn is by the so called unsupervised learning methods, where programs/agents learn without any human interaction. Psychologists and Neurologists have long tried to understand how the human brain works, but due to its complexity there are still some obstacles left before we will be able to simulate the different functionalities. This thesis is an attempt to get one step closer to solving the problem of how learning happens and memories form. If we were to be able to simulate human learning in a machine there is no telling where it could end. Jørn Hokland has put forward three learning rules that may describe how learning happens. These rules will be examined and then used in an artificial neural network with the intention to control a simulated robot. Artificial neural networks (ANNs) are more or less inspired by the biological neural networks (BNNs) found in humans and animals. As we will see this thesis seeks to be one of the more biologically inspired.

The Creek system has an architecture that facilitates combined case-based and model-based reasoning. The jColibri system, developed by the CBR group of Universidad Complutense in Madrid, contains a library of CBR system components intended for sharing and reuse. The system also contains an ontology (CBROnto) of CBR tasks and methods for explicit modelling of a CBR systems, in addition to general CBR terminology. In this master degree project, Creek and jColibri are compared with the aim of developing a mechanism for importing jColibri components to Creek, so that they can be integrated into a running Creek system. The mechanism is exemplified through selection of a few specific components, and integration of these components into an implemented demonstrator system. In addition, efforts needed to bring Creek into the jColibri framework are identified.

You can see sliding doors everywhere, be it at the grocery store or the hospital. These doors are today mostly based on naive, motion sensing, and hence not very intelligent in deciding to open or not. I propose a solution by replacing the traditional sensor with the more sophisticated Microsoft Kinect depth mapping sensor allowing for skeletal tracking and feature extraction. I have applied hidden markov models to the behavioural features to understand the human intentions. Combined with a few simple rules, this solution proved to be accurate in 4 out of 5 times in understanding the user's intention in a controlled laboratory test.

Lung cancer is one of the deadliest and most common types of cancer inNorway. Early and precise diagnosis is crucial for improving the survivalrate. Diagnosis is often done by extracting a tissue sample in the lung throughthe mouth and throat. It is difficult to navigate to the tissue because of thecomplexity of the airways inside the lung and the reduced visibility. Our goalis to make a program that can automatically extract a map of the Airwaysdirectly from X-ray Computer Tomography(CT) images of the patient. Thisis a complex task and requires time consuming processing.In this thesis we explore different methods for extracting the Airways fromCT images. We also investigate parallel processing and the usage of moderngraphic processing units for speeding up the computations. We rate severalmethods in terms of reported performance and the possibility of parallelprocessing. The best rated method is implemented in a parallel frameworkcalled Open Computing Language.The results shows that our implementation is able to extract large parts ofthe Airway Tree, but struggles with the smaller airways and airways thatdeviate from a perfect circular cross-section. Our implementation is ableto process a full CT scan using less than a minute with a modern graphicprocessing units. The implementation is very general and is able to extractother tubular structures as well. To show this we also run our implementationon a Magnetic Resonance Angio dataset for finding blood vessels in the brainand achieve good results.We see a lot of potential in this method for extracting tubular structures. Themethod struggles the most with noise handling and tubes that deviate froma circular cross-sectional shape. We believe that this can be improved byusing another method than ridge traversal for the centerline extraction step.Because this is a local greedy algorithm, it often terminates prematurely dueto noise and other image artifacts.