The tuning of optimization software is of key interest to researchers solving mixed integer programming (MIP) problems. The efficiency of the optimization software can be greatly impacted by the solver’s parameter settings and the structure of the MIP. A designed experiment approach is used to fit a statistical model that would suggest settings of the parameters that provided the largest reduction in the primal integral metric. Tuning exemplars of six and 59 factors (parameters) of optimization software, experimentation takes place on three classes of MIPs: survivable fixed telecommunication network design, a formulation of the support vector machine with the ramp loss and L1-norm regularization, and node packing for coding theory graphs. This research presents and demonstrates a framework for tuning a portfolio of MIP instances to not only obtain good parameter settings used for future instances of the same class of MIPs, but to also gain insights into which parameters and interactions of parameters are significant for that class of MIPs. The framework is used for benchmarking of solvers with tuned parameters on a portfolio of instances. A group screening method provides a way to reduce the number of factors in a design and reduces the time it takes to perform the tuning process. Portfolio benchmarking provides performance information of optimization solvers on a class with instances of a similar structure.

Aprendizado de máquina é uma área de pesquisa na qual se investiga como desenvolver sistemas capazes de aprender com a experiência. Muitos algoritmos de aprendizado possuem parâmetros cujos valores devem ser especificados pelo usuário. Em geral, esses valores influenciam diretamente no processo de aquisição do conhecimento, podendo gerar diferentes modelos. Recentemente, algoritmos de otimização bioinspirados têm sido aplicados com sucesso no ajuste de parâmetros de técnicas de aprendizado de máquina. Essas técnicas podem apresentar diferentes sensibilidades em relação aos valores escolhidos para seus parâmetros e diferentes algoritmos de ajuste de parâmetros podem apresentar desempenhos singulares. Esta dissertação investiga a utilização de algoritmos bioinspirados para o ajuste de parâmetros de redes neurais artificiais e máquinas de vetores de suporte em problemas de classificação. O objetivo dessa investigação é verificar quais são as técnicas que mais se beneficiam do ajuste de parâmetros e quais são os algoritmos mais eficientes para essas técnicas. Os resultados experimentais mostram que os algoritmos bioinspirados conseguem encontrar melhores clasificadores que outras abordagens. Porém, essa melhoria é estatisticamente significativa para alguns conjuntos de dados. Foi possível verificar que o uso dos valores padrão para os parâmetros das técnicas de classificação leva a desempenhos similares aos obtidos com os algoritmos bioinspirados. Entretanto, para alguns conjuntos de dados, o ajuste de parâmetros pode melhorar significativamente o desempenho dos classificadores / Machine learning is a research area whose main goal is to design computational systems capable of learning through experience. Many machine learning techniques have free parameters whose values are generally defined by the user. Usually, these values affect the knowledge acquisition process directly, resulting in different models. Recently, bioinspired optimization algorithms have been successfully applied to the parameter tuning of machine learning techniques. These techniques may present variable sensitivity to the selection of the values of its parameters and different parameter tuning algorithms may present different behaviors. This thesis investigates the use of bioinspired algorithms for the parameter tuning of artificial neural networks and support vector machines in classification problems. The goal of this thesis is to investigate which techniques benefits most from parameter tuning and which are the most efficient algorithms to use with these techniques. Experimental results show that these bioinspired algorithms can find better classifiers when compared to other approaches. However, this improvement is statistically significant only to some datasets. It was possible to verify that the use of standard parameter values for the classification techniques leads to similar performances to those obtained with the bioinspired algorithms. However, for some datasets, the parameter tuning may significantly improve a classifier performance

[pt] Esta dissertação tem como objetivo avaliar algoritmos de segmentação para imagens de sensoriamento remoto. Quatro algoritmos de segmentação foram considerados neste estudo. Esses algoritmos têm abordagens diferentes tais como baseado em agrupamento, em crescimento de regiões, em modelos bayesianos e em grafos. Como cada algoritmo tem os seus próprios parâmetros, o processo de encontrar seus parâmetros ótimos foi feito usando um algoritmo de otimização, Nelder - Mead. O algoritmo Nelder - Mead procura os melhores parâmetros para cada algoritmo de segmentação, isto é, os parâmetros que proporcionam os resultados mais exatos com respeito a uma referência dada. A função objetivo foi definida a partir de sete métricas diferentes. Eles avaliam qualitativamente o resultado da segmentação baseadas na sua referência. Os experimentos foram realizados ao longo de três imagens de sensoriamento remoto de diferentes localidades do Brasil. Isso envolveu um total de 84 experimentos. Os resultados mostraram que as abordagens baseadas em grafos produzem os melhores resultados baseados em todas as métricas. As abordagens baseadas no crescimento de regiões e agrupamento apresentaram-se como boas opções para imagens de sensoriamento remoto. / [en] This dissertation aims to evaluate segmentation algorithms for remote sensing images. Four segmentation algorithms were considered in this study. These algorithms have different approaches such as clustering-based, region growing-based, bayesian-based and graph-based. As each algorithm has its own parameters, the process to find their optimum values was done using an optimization algorithm, Nelder - Mead. Nelder - Mead algorithm looks for the best parameters for each segmentation algorithm, i.e. the parameters that provide the most accurate results with respect to a given reference. The objective function was defined by seven different metrics. These metrics assess qualitatively the segmentation result based on its reference. The experiments were performed over three remote sensing images from different locations of Brazil. A total of 84 experiments have been performed. The results have shown that graph-based approaches produce the best results based on each metric. The region growing- and clustering-based approaches have shown to be good alternatives for remote sensing images.

AMSR-E is a space-borne radiometer which measures Earth microwave emissions or brightness temperatures (Tb) over a wide swath. AMSR-E data and images are useful in mapping valuable Earth-surface and atmospheric phenomena. A modified version of the Scatterometer Image Reconstruction (SIR) algorithm creates Tb images from the collected data. SIR is an iterative algorithm with tuning parameters to optimize the reconstruction for the instrument and channel. It requires an approximate aperture function for each channel to be effective. This thesis presents a simulator-based optimization of SIR iteration and aperture function threshold parameters for each AMSR-E channel. A comparison of actual Tb images generated using the optimal and sub-optimal values is included. Tuned parameters produce images with sharper transitions between regions of low and high Tb for lower-frequency channels. For higher-frequency channels, the severity of artifacts due to temporal Tb variation of the input measurements decreases and coverage gaps are eliminated after tuning. A two-parameter Gaussian-like bell model is currently assumed in image reconstruction to approximate the AMSR-E aperture function. This paper presents a method of estimating the effective AMSR-E aperture function using Tb measurements and geographical information. The estimate is used as an input for image reconstruction. The resulting Tb images are compared with those produced with the previous Gaussian approximation. Results support the estimates found in this paper for channels 1h, 1v, and 2h. Images processed using the old or new aperture functions for all channels differed by a fraction of a Kelvin over spatially smooth regions.

No description available.

I April 2018 an experiment was conducted at the university of Bergen in collaboration with Novelda using the XeThru range-doppler radar. The goal of the experiment was to record data on rats using XeThru radar and by using traditional medical sensors, to see if a XeThru radar can be used instead of the traditional sensors. The data from the experiment consisted of data saved as EDF (European Data Format) from the traditional sensors, radar data and video. The rest goal of this thesis was to analyze the different data types. Then to look at the radar data with the goal of seeing if respiration detection on rats was possible and try to create a radar pro le for use on rats. This report focuses on only one of the rats from the experiment, because this rat was recorded using a video camera. In a preliminary project a python program had been written to use on the EDF data. The program was expanded upon in this thesis to optimize the performance. The python program took the EDF data and extracted the desired signal, data from a muscle sensor, and plotted it, making it possible to analyze. For the radar data Matlab was used, since Novelda already had the tools to process the data using that program. This worked using parameter les to set all the necessary settings to do a playback and change the radar and processing settings. The work then consisted of changing the parameters to get the best results on the data from the rats. A parameter le was created for use on rats that was based on the parameter le for respiration detection on adult humans. The conclusion on the EDF data was that it was not possible to nd a pattern for respiration using the muscle sensor. The radar data fortunately yielded better results. After much testing and optimization it was possible to get good respiration detection on rat 7. Some of the e ndings was possible noise problems and sources. The radar was mounted to the rack holding the cages. This meant it was susceptible to vibration noise form the other rats in the same rack and to the construction work being done at the university. Another problem with mounting of the radar was that is was to close to the cage, because the rat slept close to the wall of the cage closest to the radar. Then the problem of directly coupled energy interfered with the actual signal form the radar. It is still unclear if a radar sensor can replace the traditional sensors, but it is very possible to detect respiration on rats. With some more work, and possible a new experiment xing the noise sources, can make it possible.

No description available.

Multi-objective decision-making is critical for everyday tasks and engineering problems. Finding the perfect trade-off to maximize all the solution's criteria requires a considerable amount of experience or the availability of a significant number of resources. This makes these decisions difficult to achieve for expensive problems such as engineering. Most of the time, to solve such expensive problems, we are limited by time, resources, and available expertise. Therefore, it is desirable to simplify or approximate the problem when possible before solving it. The state-of-the-art approach for simplification is model-based or surrogate-based optimization. These approaches use approximation models of the real problem, which are cheaper to evaluate. These models, in essence, are simplified hypotheses of cause-effect relationships, and they replace high estimates with cheap approximations. In this thesis, we investigate surrogate models as wrappers for the real problem and apply \gls{moea} to find Pareto optimal decisions. The core idea of surrogate models is the combination and stacking of several models that each describe an independent objective. When combined, these independent models describe the multi-objective space and optimize this space as a single surrogate hypothesis - the surrogate compositional model. The combination of multiple models gives the potential to approximate more complicated problems and stacking of valid surrogate hypotheses speeds-up convergence. Consequently, a better result is obtained at lower costs. We combine several possible surrogate variants and use those that pass validation. After recombination of valid single objective surrogates to a multi-objective surrogate hypothesis, several instances of \gls{moea}s provide several Pareto front approximations. The modular structure of implementation allows us to avoid a static sampling plan and use self-adaptable models in a customizable portfolio. In numerous case studies, our methodology finds comparable solutions to standard NSGA2 using considerably fewer evaluations. We recommend the present approach for parameter tuning of expensive black-box functions.:1 Introduction 1.1 Motivation 1.2 Objectives 1.3 Research questions 1.4 Results overview 2 Background 2.1 Parameter tuning 2.2 Multi-objective optimization 2.2.1 Metrics for multi-objective solution 2.2.2 Solving methods 2.3 Surrogate optimization 2.3.1 Domain-specific problem 2.3.2 Initial sampling set 2.4 Discussion 3 Related Work 3.1 Comparison criteria 3.2 Platforms and frameworks 3.3 Model-based multi-objective algorithms 3.4 Scope of work 4 Compositional Surrogate 4.1 Combinations of surrogate models 4.1.1 Compositional Surrogate Model [RQ1] 4.1.2 Surrogate model portfolio [RQ2] 4.2 Sampling plan [RQ3] 4.2.1 Surrogate Validation 4.3 Discussion 5 Implementation 5.1 Compositional surrogate 5.2 Optimization orchestrator 6 Evaluation 6.1 Experimental setup 6.1.1 Optimization problems 6.1.2 Optimization search 6.1.3 Surrogate portfolio 6.1.4 Benchmark baseline 6.2 Benchmark 1: Portfolio with compositional surrogates. Dynamic sampling plan 6.3 Benchmark 2: Inner parameters 6.3.1 TutorM parameters 6.3.2 Sampling plan size 6.4 Benchmark 3: Scalability of surrogate models 6.5 Discussion of results 7 Conclusion 8 Future Work A Appendix A.1 Benchmark results on ZDT DTLZ, WFG problems

京都大学 / 0048 / 新制・課程博士 / 博士(工学) / 甲第18942号 / 工博第3984号 / 新制||工||1614(附属図書館) / 31893 / 京都大学大学院工学研究科機械理工学専攻 / (主査)教授 松野 文俊, 教授 椹木 哲夫, 教授 藤本 健治 / 学位規則第4条第1項該当 / Doctor of Philosophy (Engineering) / Kyoto University / DGAM

Achieving robust and reliable autonomous operation even in complex unstructured environments is a central goal of field robotics. As the environments and scenarios to which robots are applied have continued to grow in complexity, so has the challenge of properly defining preferences and tradeoffs between various actions and the terrains they result in traversing. These definitions and parameters encode the desired behavior of the robot; therefore their correctness is of the utmost importance. Current manual approaches to creating and adjusting these preference models and cost functions have proven to be incredibly tedious and time-consuming, while typically not producing optimal results except in the simplest of circumstances. This thesis presents the development and application of machine learning techniques that automate the construction and tuning of preference models within complex mobile robotic systems. Utilizing the framework of inverse optimal control, expert examples of robot behavior can be used to construct models that generalize demonstrated preferences and reproduce similar behavior. Novel learning from demonstration approaches are developed that offer the possibility of significantly reducing the amount of human interaction necessary to tune a system, while also improving its final performance. Techniques to account for the inevitability of noisy and imperfect demonstration are presented, along with additional methods for improving the efficiency of expert demonstration and feedback. The effectiveness of these approaches is confirmed through application to several real world domains, such as the interpretation of static and dynamic perceptual data in unstructured environments and the learning of human driving styles and maneuver preferences. Extensive testing and experimentation both in simulation and in the field with multiple mobile robotic systems provides empirical confirmation of superior autonomous performance, with less expert interaction and no hand tuning. These experiments validate the potential applicability of the developed algorithms to a large variety of future mobile robotic systems.