Catastrophic Forgetting is a behavior seen in artificial neural networks (ANNs) when new information overwrites old in such a way that the old information is no longer usable. Since this happens very rapidly in ANNs, it leads to both major practical problems and problems using the artificial networks as models for the human brain. In this thesis I will approach the problem from the practical viewpoint and attempt to provide rules, guidelines, datasets and analysis methods that can aid researchers better analyze new ANN models in terms of catastrophic forgetting and thus lead to better solutions. I suggest two methods of analysis that measure the overlap between input patterns in the input space. I will show strong indications that these measurements can predict if a back-propagation network will retain information better or worse. I will also provide source code implemented in Matlab for analyzing datasets, both with the new suggested measurements and other existing ones, and for running experiments measuring the catastrophic forgetting.

The use of video surveillance in public areas is ever increasing. With that increase, it becomes impractical to continue using humans to view and respond to the surveillance video streams, due to the massive amount of information that must be processed. If one hope to use surveillance to avoid personal injuries, damage to property and so forth, instead of merely a forensic tool after the fact, humans must be replaced by artificial intelligence. This thesis examines the whole process of recognizing unwanted human behaviors from videos taken by surveillance cameras. An overview of the state of the art in automated security and human behavior recognition is given. Algorithms for motion detection and tracking are described and implemented. The motion detection algorithm uses background subtraction, and can deal with large amounts of random noise. It also detects and removes cast shadows. The tracking algorithm uses a spatial occupancy overlap test between the predicted positions of tracked objects and current foreground blobs. Merges/splits are handled by grouping/ungrouping objects and recovering the trajectory using distance between predicted position and foreground blobs. Behaviors that are unwanted in most public areas are discussed, and a set of such concrete behaviors described. New algorithms for recognizing chasing/fleeing scenarios and people lying on the floor are then presented. A real-time intelligent surveillance system capable of recognizing chasing/fleeing scenarios and people lying on the floor has been implemented, and results from analyzing real video sequences are presented. The thesis concludes with a discussion on the advantages and disadvantes of the presented algorithms, and suggestions for future research.

Denne hovedoppgaven er rettet mot en praktisk implementasjon av et automatisk visuelt inspeksjonssystem. På bakgrunn av en serie av treningsbilder er målsetningen at systemet skal kunne klassifisere objekters avbildninger som godkjent eller underkjent. Arbeidet har lagt stor vekt på at systemet skal virke på generelle objekter. Systemet er implementert i Microsoft Visual Studio .NET 2003 C++, og viktige elementer tilknyttet arbeidet beskrives i denne rapporten. Resultatene virker lovende, da inspeksjonssystemet gjennomsnittlig klassifiserer 91% riktig på de 8 bildesett som systemet er testet med. Videre planer gir imidlertid håp om å utbedre systemet betydelig. Disse planene presenteres som videre arbeid i slutten av rapporten.

In this thesis, we develop a meta-reasoning system based on CBR which solves sliding-block puzzles. The meta-reasoning system is built on top of a search-based sliding-block puzzle solving program which was developed as part of the specialization project at NTNU. As part of the thesis work, we study existing literature on automatic puzzle solving methods and state space search, as well as the use of reasoning and meta-level reasoning applied to puzzles and games. The literature study forms the theoretical foundation for the development of the meta-reasoning system. The meta-reasoning system is further enhanced by adding a meta-control cycle which uses randomized search to generate new cases to apply to puzzles. In addition, we explore several ways of improving the underlying solver program by trying to solve hard puzzles by using the solution for easier variants, and by developing a more memory-efficient way of representing puzzle configurations. We evaluate the results of our system, and shows that it offers a slight improvement compared to solving the puzzles with a set of general cases, as well as showing vast improvement for a few isolated test cases, but the performance is slightly behind the hand-tuned parameters we found in the specialization project. We conclude our work by identifying parts of our system where improvement can be done, as well as suggesting other promising areas for further research.

This report will explore behavior-based robotics and relevant AI techniques. A system for autonomous control of mobile robots inspired by behavior-based robotics, in particular Rodney Brooks' subsumption architecture, have been implemented, adapted for use in a multiagent environment. The system is modular and flexible, allowing for easy addition and removal of system parts. A weight-based command fusion approach is taken to action selection, making it possible to satisfy multiple behaviors simultaneously.

This project report describes a hybrid genetic algorithm that works as a schedule generator for a complex robotic harvesting task. The task is set to a dynamic environment with a robotic opponent, making responsiveness of the planning algorithm particularly important. To solve this task, many previous scheduling algorithms were studied. Genetic algorithms have successfully been used in many dynamic scheduling tasks, due to their ability to incrementally adapt and optimize solutions when changes are made to the environment. Many of the previous approaches also used a separate heuristic to quicly adapt solutions to the new environment, making the algorithm more responsive. In addition, the study of previous work revealed the importance of population diversity when making a responsive genetic algorithm. Implementation was based on a genetic algorithm made as the author's fifth year specialization project for solving a static version of the same task. This algorithm was hybridized with a powerful local search technique that proved essential in generating good solutions for the complex harvesting task. When extending the algorithm to also work in a dynamically changing environment, several adaptations and extensions needed to be made, to make it more responsive. The extensions and adaptations include a fast-response heuristic for immediate adaptation to environmental changes, a decrease in genotype size to speed up local searches and a contingency planning module intending to solve problems before they arise. Experiments proved that the implemented dynamic planner successfully adapted its plans to a changing environment, clearly showing improvements compared to running a static plan. Further experiments also proved that the dynamic planner was able to deal with erroneous time estimates in its simulator module in a good way. Experiments on contingency planning gave no clear results, but indicated that using computational resources for planning ahead may be a good choice, if the contingency to plan for is carefully selected. As no unequivocal results were obtained, further studies of combining genetic algorithms and contingency planning may be an interesting task for future efforts.

A summary is a shortened version of a text that contains the main points of the original content. Automatic summarization is the task of generating a summary by a computer. For example, given a collection of news articles for the last week an automatic summarizer is able to create a concise overview of the important events. This summary can be used as the replacement for the original content or help to identify the events that a person is particularly interested in. Potentially, automatic summarization can save a lot of time for people that deal with a large amount of textual information. The straightforward way to generate a summary is to select several sentences from the original text and organize them in way to create a coherent text. This approach is called extraction-based summarization and is the topic of this thesis. Extraction-based summarization is a complex task that consists of several challenging subtasks. The essential part of the extraction-based approach is identification of sentences that contain important information. It can be done using graph-based representations and centrality measures that exploit similarities between sentences to identify the most central sentences. This thesis provide a comprehensive overview of methods used in extraction-based automatic summarization. In addition, several general natural language processing issues such as feature selection and text representation models are discussed with regard to automatic summarization. Part of the thesis is dedicated to graph-based representations and centrality measures used in extraction-based summarization. Theoretical analysis is reinforced with the experiments using the summarization framework implemented for this thesis. The task for the experiments is query-focused multi-document extraction-based summarization, that is, summarization of several documents according to a user query. The experiments investigate several approaches to this task as well as the use of different representation models, similarity and centrality measures. The obtained results indicate that use of graph centrality measures significantly improves the quality of generated summaries. Among the variety of centrality measure the degree-based ones perform better than path-based measures. The best performance is achieved when centralities are combined with redundancy removal techniques that prevent inclusion of similar sentences in a summary. Experiments with representation models reveal that a simple local term count representation performs better than the distributed representation based on latent semantic analysis, which indicates that further investigation of distributed representations in regard to automatic summarization is necessary. The implemented system performs quite good compared with the systems that participated in DUC 2007 summarization competition. Nevertheless, manual inspection of the generated summaries demonstrate some of the flaws of the implemented summarization mechanism that can be addressed by introducing advanced algorithms for sentence simplification and sentence ordering.

This thesis investigates the application of artificial neural networks (ANNs) for forecasting financial time series (e.g. stock prices).The theory of technical analysis dictates that there are repeating patterns that occur in the historic prices of stocks, and that identifying these patterns can be of help in forecasting future price developments. A system was therefore developed which contains several ``agents'', each producing recommendations on the stock price based on some aspect of technical analysis theory. It was then tested if ANNs, using these recommendations as inputs, could be trained to forecast stock price fluctuations with some degree of precision and reliability.The predictions of the ANNs were evaluated by calculating the Pearson correlation between the predicted and actual price changes, and the ``hit rate'' (how often the predicted and the actual change had the same sign). Although somewhat mixed overall, the empirical results seem to indicate that at least some of the ANNs were able to learn enough useful features to have significant predictive power. Tests were performed with ANNs forecasting over different time frames, including intraday. The predictive performance was seen to decline on the shorter time scales.

In this thesis we develop a novel network model that extends the traditional artificial neural network (ANN) model to include oscillatory behaviour. This model is able to correctly classify combinations of previously learned input patterns by grouping features that belong to the same category. This grouping process is termed segmentation and we show how synchrony of oscillations is the necessary missing component of ANNs to be able to perform this segmentation. Using this model we go on to show that top-down modulatory feedback is necessary to enable separation of multiple objects in a scene and segmentation of their individual features. This type of feedback is distinctly different than recurrency and is what enables the rich dynamics between the nodes of our network. Additionally, we show how our model's dynamics avoid the combinatorial explosion in required training repetitions of traditional feed-forward classification networks. In these networks, relations between objects must explicitly be learned. In contrast, the dynamics of modulatory feedback allow us to defer calculation of these relations until run-time, thus creating a more robust system. We call our model Fuzzy Oscillations, and it achieves good results when compared to existing models. However, oscillatory neural network models successful in achieving segmentation are a relatively recent development. We thus feel that our model is a contribution to the field of oscillatory neural networks.

Historical stock prices are used to predict the direction of future stock prices. The developed stock price prediction model uses a novel two-layer reasoning approach that employs domain knowledge from technical analysis in the first layer of reasoning to guide a second layer of reasoning based on machine learning. The model is supplemented by a money management strategy that use the historical success of predictions made by the model to determine the amount of capital to invest on future predictions. Based on a number of portfolio simulations with trade signals generated by the model, we conclude that the prediction model successfully outperforms the Oslo Benchmark Index (OSEBX).