Scholarship & Research
Permanent URI for this communityhttps://scholarworks.montana.edu/handle/1/1
Browse
5 results
Search Results
Item Factored evolutionary algorithms: cooperative coevolutionary optimization with overlap(Montana State University - Bozeman, College of Engineering, 2017) Strasser, Shane Tyler; Chairperson, Graduate Committee: John SheppardFactored Evolutionary Algorithms (FEA) define a relatively new class of evolutionary-based optimization algorithms that have been successfully applied to various problems, such as training neural networks and performing abductive inference in graphical models. FEA is unique in that it factors the function being optimized by creating subpopulations that optimize over a subset of dimensions of the function. However, unlike other optimization techniques that subdivide optimization problems, FEA encourages subpopulations to overlap with one another, allowing subpopulations to compete and share information. Although FEA has been shown to be very effective at function optimization, there is still little understanding with respect to its general characteristics. In this dissertation, we present seven results exploring the theoretical and empirical properties of FEA. First, we present a formal definition of FEA and demonstrate its relationships to other multiple population algorithms. Second, we demonstrate that FEA's success is independent of the underlying optimization algorithm by evaluating the performance of FEA using a wide variety of evolutionary- and swarm-based algorithms over single-population and non-overlapping versions. Third, we demonstrate that for a given problem, there is an optimal way to generate groups of overlapping subpopulations derived using the Markov blanket in Bayesian networks. Fourth, we establish that a class of optimization functions like NK landscapes can be mapped directly to probabilistic graphical models. Additionally, we demonstrate that factor architectures derived from Markov blankets maintain better diversity of individuals in their population. Fifth, we present a new discrete Particle Swarm Optimization (PSO) algorithm and compare its performance to competing approaches. In addition, we analyze the performance of FEA versions of discrete PSO and discover that FEA masks the poor performance of search algorithms. We show what conditions are necessary for FEA to converge and scenarios where FEA may become stuck in suboptimal regions in the search space. Finally, we explore the performance of FEA on unitation functions and discover several instances where FEA struggles to outperform single-population algorithms. These results allow us to determine which situations are appropriate for FEA when using solving real-world problems.Item Learning spectral filters for single- and multi-label classification of musical instruments(Montana State University - Bozeman, College of Engineering, 2015) Donnelly, Patrick Joseph; Chairperson, Graduate Committee: John SheppardMusical instrument recognition is an important research task in music information retrieval. While many studies have explored the recognition of individual instruments, the field has only recently begun to explore the more difficult multi-label classification problem of identifying the musical instruments present in mixtures. This dissertation presents a novel method for feature extraction in multi-label instrument classification and makes important contributions to the domain of instrument classification and to the research area of multi-label classification. In this work, we consider the largest collection of instrument samples in the literature. We examine 13 musical instruments common to four datasets. We consider multiple performers, multiple dynamic levels, and all possible musical pitches within the range of the instruments. To the area of multi-label classification, we introduce a binary-relevance feature extraction scheme to couple with the common binary-relevance classification paradigm, allowing selection of features unique to each class label. We present a data-driven approach to learning areas of spectral prominence for each instrument and use these locations to guide our binary-relevance feature extraction. We use this approach to estimate source separation of our polyphonic mixtures. We contribute the largest study of single- and multi-label classification in musical instrument literature and demonstrate that our results track with or improve upon the results of comparable approaches. In our solo instrument classification experiments, we provide the seminal use of Bayesian classifiers in the domain and demonstrate the utility of conditional dependencies between frequency- and time-based features for the instrument classification problem. For multi-label instrument classification, we explore the question of dataset bias in a cross-validation study controlled for dataset independence. Additionally, we present a comprehensive cross-dataset study and demonstrate the generalizability of our approach. We consider the difficulty of the multi-label problem with regards to label density and cardinality and present experiments with a reduced label set, comparable to many studies in the literature, and demonstrate the efficacy of our system on this easier problem. Furthermore, we provide a comprehensive set of multi-label evaluation measures.Item Inference and learning in Bayesian networks using overlapping swarm intelligence(Montana State University - Bozeman, College of Engineering, 2015) Fortier, Nathan Lee; Chairperson, Graduate Committee: John SheppardWhile Bayesian networks provide a useful tool for reasoning under uncertainty, learning the structure of these networks and performing inference over them is NP-Hard. We propose several heuristic algorithms to address the problems of inference, structure learning, and parameter estimation in Bayesian networks. The proposed algorithms are based on Overlapping Swarm intelligence, a modification of particle swarm optimization in which a problem is broken into overlapping subproblems and a swarm is assigned to each subproblem. The algorithm maintains a global solution that is used for fitness evaluation, and is updated periodically through a competition mechanism. We describe how the problems of inference, structure learning, and parameter estimation can be broken into subproblems, and provide communication and competition mechanisms that allow swarms to share information about learned solutions. We also present a distributed alternative to Overlapping Swarm Intelligence that does not require a global network for fitness evaluation. For the problems of full and partial abductive inference, a swarm is assigned to each relevant node in the network. Each swarm learns the relevant state assignments associated with the Markov blanket for its corresponding node. In our approach to parameter estimation, a swarm is associated with each node in the network that corresponds to either a latent variable or a child of a latent variable. Each node's corresponding swarm learns the parameters associated with that node's Markov blanket. We also apply Overlapping Swarm Intelligence to several variations of the structure learning problem: learning Bayesian classifiers, learning Bayesian networks with complete data, and learning Bayesian networks with latent variables. For each problem, a swarm is associated with each node in the network. This work makes a number of contributions relating to the advancement of Overlapping Swarm Intelligence as a general optimization technique. We demonstrate the applicability of Overlapping Swarm Intelligence to both discrete and continuous optimization problems. We also examine the effect of the swarm architecture and degree of overlap on algorithm performance. The experiments presented here demonstrate that, while the sub-swarm architecture affects algorithm performance, Overlapping Swarm Intelligence continues to perform well even when there is little overlap between the swarms.Item Netoracle, an intelligent information agent(Montana State University - Bozeman, College of Engineering, 1998) Van den Bogaert, JorisSearching the World Wide Web today is comparable to giving a librarian a list of words and expecting him to find a set of documents that exactly answer your request. Because so little information is supplied, the librarian can only return partially satisfactory results. For example, some returned documents might be irrelevant or not enough relevant documents are found. There is no easy solution to producing a concise, valuable response to a query due to the enormous amount of information on the Web combined with the lack of homogeneity among Web pages. This thesis addresses this problem and provides a prototype of a search tool that exhibits some intelligent behavior. A number of improvements to the current state of the art in search engines are suggested.Item A system for automating identification of biological echoes in NEXRAD level II radar data(Montana State University - Bozeman, College of Engineering, 2009) Mead, Reginald Marshall; Chairperson, Graduate Committee: John PaxtonSince its inception in the mid twentieth century, radar ornithology has provided scientists with new tools for studying the behavior of birds, especially with regards to migration. A number of studies have shown that birds can be detected using a wide variety of radar devices. Generally, these studies have focused on small portable radars that typically have a finer resolution than large weather surveillance radars. Recently, however, a number of researchers have presented qualitative evidence suggesting that birds, or at least migration events, can be identified using large broad scale radars such as the WSR-88D used in the NEXRAD weather surveillance system. This is potentially a boon for ornithologists because NEXRAD data covers a large portion of the country, is constantly being produced, is freely available, and is archived back into the early 1990s. A major obstacle is that identifying birds in NEXRAD data currently requires having a trained technician manually inspect a graphically rendered radar sweep. The immense amount of available data makes manual classification of radar echoes infeasible over any practical span of space or time. In this thesis, a system is presented for automating this process using machine learning techniques. This approach begins with classified training data that has been interpreted by experts or collected from direct observations. The data is preprocessed to ensure quality and to emphasize relevant features. A classifier is then trained using this data and cross validation is used to measure performance. The experiments in this thesis compare neural network, naïve Bayes, and k-nearest neighbor classifiers. Empirical evidence is provided showing that this system can achieve classification accuracies in the 80th to 90th percentile.