Most of the existing poker agents using case-based reasoning (CBR) are based on imitation of other poker agents and have weak capabilities of adapting their own strategies to different opponents or playing styles. We address these concerns in the development of UpperCase, a heads up no-limit Texas Hold'em poker agent representing a new approach to the application of CBR in poker. Using methods of perfect information hindsight analysis, the poker agent attempts to more accurately determine the quality of poker decisions. Through extensive exploration of the quality of different decisions, UpperCase is able to invent new poker strategies. The agent also tries to recognize different opponents by observing their actions and perform adaptation accordingly. Experimental results suggest that the agent is able to successfully create new profitable strategies, as well as achieve increased performance by dynamically changing its strategy during play.

This thesis concerns the modeling of running speed in orient-eering by means of multi-layered feed forward artificial neu-ral networks with Backpropagation learning, using GPS tracks of orienteers, an orienteering map and a digital eleva-tion model as the basis for the training data. A learning sys-tem was implemented and tested with GPS data collected by a test subject. A trained speed model was applied in a third-party application for arithmetic analysis of route choices. The proposed method is shown to have potential, even though the results as of now are not good enough to be considered useful to orienteers.

Periods of sub-optimal production rates, or complete shut-downs, add negative numbers to the revenuegraph for oil companies. Oil and gas are produced from several reservoirs and through many wells withvarying gas/oil proportion, making it a complex process that is difficult to control. As a part of a threestepprocess for utilizing data in the oil production domain, this thesis derive methods for combiningevent patterns, called restricted association rules, in time series in order to warn about future anomalies inoil production processes. Two problems have been considered: Network learning and network reasoning.The suggested solution consists of building an Association Rules Network (ARN) from the rule set givenas input. After transforming the hypergraph-based ARN to a directed acyclic graph, correlations betweennodes are found by applying the shortest-path principle. Motivated by the shortcomings of this simplesolution, it is shown how a method for learning Bayesian networks with support for representation oftemporal dependencies can be derived from the initial ARN. The concept, named Temporal BayesianNetwork of Events (TBNE), is a powerful, but yet complex solution that enjoys the properties of Bayesiannetwork reasoning while at the same time representing temporal information. This thesis has shown thatit is theoretically feasible to combine restricted association rules in order to create a network structurefor reasoning. It is concluded that the final choice of solution must be based on a carefully considerationof the trade-off between complexity and expressiveness, and that a natural continuation is testing thesuggested concepts with real data.

Periods of sub-optimal production rates, or complete shut-downs, add negative numbers to the revenue graph for oil companies. Oil and gas are produced from several reservoirs and through many wells with varying gas/oil proportion, making it a complex process that is difficult to control. As a part of a three step process for utilizing data in the oil production domain, this thesis derive methods for discovering event patterns, called restricted association rules, from time series in order to pre-warn about future problems in oil production processes. A restricted rule syntax and semantics is derived to explicitly target rules suited for prediction. Based on the defined rule syntax, a two step process is derived where restricted rule mining based on the concept of minimal occurrences is used to discover restricted association rules from a sequence of events. Next, redundant rules are removed based on the concept of minimum improvement and chaining of rules, during a rule selection phase. Information theory is applied in order to identify the most interesting rules, which can be submitted to an expert for validation. Both a simple solution for easy implementation in ConocoPhillips and a more advanced solution appropriate for general prediction cases are derived. This thesis concludes that it is feasible to discover dependencies between events from actual process data. It is also concluded that a large number of rules can be pruned, in order to get a manageable set of rules which is believed to have good predictive performance.

The theory of Connectology sets forth three psychologically founded synaptic learning mechanisms that may describe all aspects of animal learning. Of particular interest to this thesis is the learning of animal motion behavior, or, more specifically, the development of synchronized and repetitive movement patterns - gaits.Computer simulations are performed according to the methodology of computational neuroethology: Artificial neural networks are simulated operating in a tight feedback loop with a structurally simple but mechanically realistic body and a physically realistic environment. Neural network learning is purely synaptical and is performed solely within the lifetime of one such ANN-controlled system. Additionally, the configuration parameter space is searched by means of genetic algorithms.Simulation results show examples of synchronized and repetitive movement patterns developing when neuronal and mechanical model parameters are appropriately specified. These simulations thereby provide the first examples known to us of a fully unsupervised and self-organized artificial neural system that synaptically learns synchronized and repetitive motor control. In spite of limited mechanical model complexity, the most efficient movement patterns to some degree resemble the gaits seen in nature.

Fish farming is a million dollar business world wide, and fish is in fact the third mostimportant export product after oil/gas and metal in Norway. There are a lot of different aquaculture sites which produce fish along our long coast line and they all have somedifferences in the production rates and procedures. The fish farmer at these sites holdvaluable information about the production, which is almost impossible to derive onlyfrom empirical data.In this thesis we introduce Glaucus, a Case-Based Reasoning system which aims tohelp the fish farmers with their decision making when conduction sorting operations attheir aquaculture sites. The system is built in Java and uses the jColibri developmentframework for Case-Based Reasoning. It retrieves cases based on similarity function frommyCBR and jColibri in addition to custom made ones. The case base is generated fromreal world data and the case queries are populated by a combination of user input anddata from a database with continuous data flow.Our approach is just the beginning of what we hope will be a even greater journeytowards a complete decision support system that will meet the expectations of the fishfarmers.Keywords: Case-Based Reasoning, Machine learning, Fish farming, jColibri, myCBR

The purpose of this thesis was to examine ways to adapt common 2D segmentation techniques to work with 3D grids. The focus of the thesis became how to automate and improve the performance of region growing in 3D grids. After examining relevant literature and developing a tool to run experiments, a simple automatic region grower for 3D grids was developed. Quantitative performance measures and qualitative analysis of the segmentation results were performed. This algorithm was then used as a baseline for comparison when developing a more advanced region grower for 3D grids based on the seeded region grower (SRG) for 2D grids. This new algorithm was then modified to improve its speed and later extended to allow fully automatic operation by automating the placement of starting seeds. It was found that for the algorithms that were extended to a 3D grid, the main challenge was the resources needed by these algorithms when operating on high resolution grids. It was found that even though there have been steady and rapid improvements in consumer hardware since the original region growing algorithms were used on 2D grids, the very large amounts of data resulting from an extension from surface grids to volume grids requires that special attention is paid to handling resources effectively. It was further revealed that what was considered the best data structures and algorithms for the SRG algorithm when it was first introduced, is not necessarily the best choice on todays computing hardware. Also, the conclusion is drawn that with regards to performance, it is now possible to segment volumes approximately as fast as surfaces were segmented in the early 1990s.

Evolutionary artificial neural networks can adapt to new circumstances, and handle slight changes without catastrophic failure. However, under constantly changing circumstances, resulting in unpredictable grounds for evaluating success, the lack of memory of previous adaptations are a limiting factor. While further evolution can allow adaptations to new changes, the same is required for a return to a previous environment. To reduce the need for further evolution to deal with previously seen problems, this thesis looks at an approach to encourage previous knowledge to be retained across generations. It does this using back propagation in conjunction with an implementation of the HyperNEAT neuroevolutionary algorithm.

Many problems in the field of automatic video surveillance exists today. Some have yet to be overcome. One of these problems is how a computer system automatically can determine if a situation should cause an alarm or not. To resolve this problem, the use of Case-based reasoning (CBR) is proposed. CBR is a technique that allows a system to reason about different situations and to learn from them. The aim is to produce a system that utilizes these abilities. The system should learn to recognize the situations that causes different alarms. When a situation is recognized and categorized, these false alarms can be completely avoided. This master thesis explains and shows the advantages of using such a system together with advanced image processing techniques.

The use of computer-based decision support systems within the field of health science has over the last decades been extensively researched and tested, both in controlled environments and in clinical practice. Despite the obvious benefits of utilizing such systems in the day-to-day activities, many of the designed systems fail to make the impact one could hope to achieve. We have designed and implemented a prototype of a decision support system which use both Case-Based Reasoning and probabilistic inference through a Bayesian Network as a basis for the solution. To achieve user acceptance an explanation module has been implemented which gives the user full access to the data which has been used in the reasoning process, both from the Case-Based Reasoning and the Bayesian Network. The system has shown promising results within the domain of wine recommendation, with a very high accuracy despite uncertain accuracy of the knowledge within the system. Furthermore the explanations presented to an expert conformed to the causal way of reasoning used by said expert, and was accepted as a very useful tool to get pointed in the right direction for evaluation of the solution.