Header logo is


2020


Bayesian Optimization in Robot Learning - Automatic Controller Tuning and Sample-Efficient Methods
Bayesian Optimization in Robot Learning - Automatic Controller Tuning and Sample-Efficient Methods

Marco-Valle, A.

University of Tübingen, June 2020 (thesis)

Abstract
The problem of designing controllers to regulate dynamical systems has been studied by engineers during the past millennia. Ever since, suboptimal performance lingers in many closed loops as an unavoidable side effect of manually tuning the parameters of the controllers. Nowadays, industrial settings remain skeptic about data-driven methods that allow one to automatically learn controller parameters. In the context of robotics, machine learning (ML) keeps growing its influence on increasing autonomy and adaptability, for example to aid automating controller tuning. However, data-hungry ML methods, such as standard reinforcement learning, require a large number of experimental samples, prohibitive in robotics, as hardware can deteriorate and break. This brings about the following question: Can manual controller tuning, in robotics, be automated by using data-efficient machine learning techniques? In this thesis, we tackle the question above by exploring Bayesian optimization (BO), a data-efficient ML framework, to buffer the human effort and side effects of manual controller tuning, while retaining a low number of experimental samples. We focus this work in the context of robotic systems, providing thorough theoretical results that aim to increase data-efficiency, as well as demonstrations in real robots. Specifically, we present four main contributions. We first consider using BO to replace manual tuning in robotic platforms. To this end, we parametrize the design weights of a linear quadratic regulator (LQR) and learn its parameters using an information-efficient BO algorithm. Such algorithm uses Gaussian processes (GPs) to model the unknown performance objective. The GP model is used by BO to suggest controller parameters that are expected to increment the information about the optimal parameters, measured as a gain in entropy. The resulting “automatic LQR tuning” framework is demonstrated on two robotic platforms: A robot arm balancing an inverted pole and a humanoid robot performing a squatting task. In both cases, an existing controller is automatically improved in a handful of experiments without human intervention. BO compensates for data scarcity by means of the GP, which is a probabilistic model that encodes prior assumptions about the unknown performance objective. Usually, incorrect or non-informed assumptions have negative consequences, such as higher number of robot experiments, poor tuning performance or reduced sample-efficiency. The second to fourth contributions presented herein attempt to alleviate this issue. The second contribution proposes to include the robot simulator into the learning loop as an additional information source for automatic controller tuning. While doing a real robot experiment generally entails high associated costs (e.g., require preparation and take time), simulations are cheaper to obtain (e.g., they can be computed faster). However, because the simulator is an imperfect model of the robot, its information is biased and could have negative repercussions in the learning performance. To address this problem, we propose “simu-vs-real”, a principled multi-fidelity BO algorithm that trades off cheap, but inaccurate information from simulations with expensive and accurate physical experiments in a cost-effective manner. The resulting algorithm is demonstrated on a cart-pole system, where simulations and real experiments are alternated, thus sparing many real evaluations. The third contribution explores how to adequate the expressiveness of the probabilistic prior to the control problem at hand. To this end, the mathematical structure of LQR controllers is leveraged and embedded into the GP, by means of the kernel function. Specifically, we propose two different “LQR kernel” designs that retain the flexibility of Bayesian nonparametric learning. Simulated results indicate that the LQR kernel yields superior performance than non-informed kernel choices when used for controller learning with BO. Finally, the fourth contribution specifically addresses the problem of handling controller failures, which are typically unavoidable in practice while learning from data, specially if non-conservative solutions are expected. Although controller failures are generally problematic (e.g., the robot has to be emergency-stopped), they are also a rich information source about what should be avoided. We propose “failures-aware excursion search”, a novel algorithm for Bayesian optimization under black-box constraints, where failures are limited in number. Our results in numerical benchmarks indicate that by allowing a confined number of failures, better optima are revealed as compared with state-of-the-art methods. The first contribution of this thesis, “automatic LQR tuning”, lies among the first on applying BO to real robots. While it demonstrated automatic controller learning from few experimental samples, it also revealed several important challenges, such as the need of higher sample-efficiency, which opened relevant research directions that we addressed through several methodological contributions. Summarizing, we proposed “simu-vs-real”, a novel BO algorithm that includes the simulator as an additional information source, an “LQR kernel” design that learns faster than standard choices and “failures-aware excursion search”, a new BO algorithm for constrained black-box optimization problems, where the number of failures is limited.

ics

Repository (Universitätsbibliothek) - University of Tübingen PDF DOI [BibTex]


no image
Advances in Latent Variable and Causal Models

Rubenstein, P.

University of Cambridge, UK, 2020, (Cambridge-Tuebingen-Fellowship) (phdthesis)

ei

[BibTex]

[BibTex]

2013


no image
Camera-specific Image Denoising

Schober, M.

Eberhard Karls Universität Tübingen, Germany, October 2013 (diplomathesis)

ei pn

PDF [BibTex]

2013


PDF [BibTex]


Statistics on Manifolds with Applications to Modeling Shape Deformations
Statistics on Manifolds with Applications to Modeling Shape Deformations

Freifeld, O.

Brown University, August 2013 (phdthesis)

Abstract
Statistical models of non-rigid deformable shape have wide application in many fi elds, including computer vision, computer graphics, and biometry. We show that shape deformations are well represented through nonlinear manifolds that are also matrix Lie groups. These pattern-theoretic representations lead to several advantages over other alternatives, including a principled measure of shape dissimilarity and a natural way to compose deformations. Moreover, they enable building models using statistics on manifolds. Consequently, such models are superior to those based on Euclidean representations. We demonstrate this by modeling 2D and 3D human body shape. Shape deformations are only one example of manifold-valued data. More generally, in many computer-vision and machine-learning problems, nonlinear manifold representations arise naturally and provide a powerful alternative to Euclidean representations. Statistics is traditionally concerned with data in a Euclidean space, relying on the linear structure and the distances associated with such a space; this renders it inappropriate for nonlinear spaces. Statistics can, however, be generalized to nonlinear manifolds. Moreover, by respecting the underlying geometry, the statistical models result in not only more e ffective analysis but also consistent synthesis. We go beyond previous work on statistics on manifolds by showing how, even on these curved spaces, problems related to modeling a class from scarce data can be dealt with by leveraging information from related classes residing in di fferent regions of the space. We show the usefulness of our approach with 3D shape deformations. To summarize our main contributions: 1) We de fine a new 2D articulated model -- more expressive than traditional ones -- of deformable human shape that factors body-shape, pose, and camera variations. Its high realism is obtained from training data generated from a detailed 3D model. 2) We defi ne a new manifold-based representation of 3D shape deformations that yields statistical deformable-template models that are better than the current state-of-the- art. 3) We generalize a transfer learning idea from Euclidean spaces to Riemannian manifolds. This work demonstrates the value of modeling manifold-valued data and their statistics explicitly on the manifold. Specifi cally, the methods here provide new tools for shape analysis.

ps

pdf Project Page [BibTex]


no image
Modelling and Learning Approaches to Image Denoising

Burger, HC.

Eberhard Karls Universität Tübingen, Germany, 2013 (phdthesis)

ei

[BibTex]

[BibTex]


no image
Linear mixed models for genome-wide association studies

Lippert, C.

University of Tübingen, Germany, 2013 (phdthesis)

ei

[BibTex]

[BibTex]


no image
Modeling and Learning Complex Motor Tasks: A case study on Robot Table Tennis

Mülling, K.

Technical University Darmstadt, Germany, 2013 (phdthesis)

ei

[BibTex]

[BibTex]


no image
Intention Inference and Decision Making with Hierarchical Gaussian Process Dynamics Models

Wang, Z.

Technical University Darmstadt, Germany, 2013 (phdthesis)

ei

[BibTex]

2012


Virtual Human Bodies with Clothing and Hair: From Images to Animation
Virtual Human Bodies with Clothing and Hair: From Images to Animation

Guan, P.

Brown University, Department of Computer Science, December 2012 (phdthesis)

ps

pdf [BibTex]

2012


pdf [BibTex]


no image
Scalable graph kernels

Shervashidze, N.

Eberhard Karls Universität Tübingen, Germany, October 2012 (phdthesis)

ei

Web [BibTex]

Web [BibTex]


From Pixels to Layers: Joint Motion Estimation and Segmentation
From Pixels to Layers: Joint Motion Estimation and Segmentation

Sun, D.

Brown University, Department of Computer Science, July 2012 (phdthesis)

ps

pdf [BibTex]

pdf [BibTex]


no image
Learning Motor Skills: From Algorithms to Robot Experiments

Kober, J.

Technische Universität Darmstadt, Germany, March 2012 (phdthesis)

ei

PDF [BibTex]

PDF [BibTex]


no image
Structure and Dynamics of Diffusion Networks

Gomez Rodriguez, M.

Department of Electrical Engineering, Stanford University, 2012 (phdthesis)

ei

Web [BibTex]

Web [BibTex]


no image
Blind Deconvolution in Scientific Imaging & Computational Photography

Hirsch, M.

Eberhard Karls Universität Tübingen, Germany, 2012 (phdthesis)

ei

Web [BibTex]

Web [BibTex]

2009


no image
Kernel Learning Approaches for Image Classification

Gehler, PV.

Biologische Kybernetik, Universität des Saarlandes, Saarbrücken, Germany, October 2009 (phdthesis)

Abstract
This thesis extends the use of kernel learning techniques to specific problems of image classification. Kernel learning is a paradigm in the field of machine learning that generalizes the use of inner products to compute similarities between arbitrary objects. In image classification one aims to separate images based on their visual content. We address two important problems that arise in this context: learning with weak label information and combination of heterogeneous data sources. The contributions we report on are not unique to image classification, and apply to a more general class of problems. We study the problem of learning with label ambiguity in the multiple instance learning framework. We discuss several different image classification scenarios that arise in this context and argue that the standard multiple instance learning requires a more detailed disambiguation. Finally we review kernel learning approaches proposed for this problem and derive a more efficient algorithm to solve them. The multiple kernel learning framework is an approach to automatically select kernel parameters. We extend it to its infinite limit and present an algorithm to solve the resulting problem. This result is then applied in two directions. We show how to learn kernels that adapt to the special structure of images. Finally we compare different ways of combining image features for object classification and present significant improvements compared to previous methods.

ei

PDF [BibTex]

2009


PDF [BibTex]


no image
Kernel Methods in Computer Vision:Object Localization, Clustering,and Taxonomy Discovery

Blaschko, MB.

Biologische Kybernetik, Technische Universität Berlin, Berlin, Germany, March 2009 (phdthesis)

ei

PDF PDF [BibTex]

PDF PDF [BibTex]


no image
Motor Control and Learning in Table Tennis

Mülling, K.

Eberhard Karls Universität Tübingen, Gerrmany, 2009 (diplomathesis)

ei

[BibTex]

[BibTex]


no image
Hierarchical Clustering and Density Estimation Based on k-nearest-neighbor graphs

Drewe, P.

Eberhard Karls Universität Tübingen, Germany, 2009 (diplomathesis)

ei

[BibTex]

[BibTex]


no image
Learning with Structured Data: Applications to Computer Vision

Nowozin, S.

Technische Universität Berlin, Germany, 2009 (phdthesis)

ei

PDF [BibTex]

PDF [BibTex]


no image
From Differential Equations to Differential Geometry: Aspects of Regularisation in Machine Learning

Steinke, F.

Universität des Saarlandes, Saarbrücken, Germany, 2009 (phdthesis)

ei

PDF [BibTex]

2005


no image
Extension to Kernel Dependency Estimation with Applications to Robotics

BakIr, G.

Biologische Kybernetik, Technische Universität Berlin, Berlin, November 2005 (phdthesis)

Abstract
Kernel Dependency Estimation(KDE) is a novel technique which was designed to learn mappings between sets without making assumptions on the type of the involved input and output data. It learns the mapping in two stages. In a first step, it tries to estimate coordinates of a feature space representation of elements of the set by solving a high dimensional multivariate regression problem in feature space. Following this, it tries to reconstruct the original representation given the estimated coordinates. This thesis introduces various algorithmic extensions to both stages in KDE. One of the contributions of this thesis is to propose a novel linear regression algorithm that explores low-dimensional subspaces during learning. Furthermore various existing strategies for reconstructing patterns from feature maps involved in KDE are discussed and novel pre-image techniques are introduced. In particular, pre-image techniques for data-types that are of discrete nature such as graphs and strings are investigated. KDE is then explored in the context of robot pose imitation where the input is a an image with a human operator and the output is the robot articulated variables. Thus, using KDE, robot pose imitation is formulated as a regression problem.

ei

PDF PDF [BibTex]

2005


PDF PDF [BibTex]


no image
Geometrical aspects of statistical learning theory

Hein, M.

Biologische Kybernetik, Darmstadt, Darmstadt, November 2005 (phdthesis)

ei

PDF [BibTex]

PDF [BibTex]


no image
Implicit Surfaces For Modelling Human Heads

Steinke, F.

Biologische Kybernetik, Eberhard-Karls-Universität, Tübingen, September 2005 (diplomathesis)

ei

[BibTex]

[BibTex]


no image
Machine Learning Methods for Brain-Computer Interdaces

Lal, TN.

Biologische Kybernetik, University of Darmstadt, September 2005 (phdthesis)

ei

Web [BibTex]

Web [BibTex]


no image
Efficient Adaptive Sampling of the Psychometric Function by Maximizing Information Gain

Tanner, TG.

Biologische Kybernetik, Eberhard-Karls University Tübingen, Tübingen, Germany, May 2005 (diplomathesis)

Abstract
A common task in psychophysics is to measure the psychometric function. A psychometric function can be described by its shape and four parameters: offset or threshold, slope or width, false alarm rate or chance level and miss or lapse rate. Depending on the parameters of interest some points on the psychometric function may be more informative than others. Adaptive methods attempt to place trials on the most informative points based on the data collected in previous trials. A new Bayesian adaptive psychometric method placing trials by minimising the expected entropy of the posterior probabilty dis- tribution over a set of possible stimuli is introduced. The method is more flexible, faster and at least as efficient as the established method (Kontsevich and Tyler, 1999). Comparably accurate (2dB) threshold and slope estimates can be obtained after about 30 and 500 trials, respectively. By using a dynamic termination criterion the efficiency can be further improved. The method can be applied to all experimental designs including yes/no designs and allows acquisition of any set of free parameters. By weighting the importance of parameters one can include nuisance parameters and adjust the relative expected errors. Use of nuisance parameters may lead to more accurate estimates than assuming a guessed fixed value. Block designs are supported and do not harm the performance if a sufficient number of trials are performed. The method was evaluated by computer simulations in which the role of parametric assumptions, its robustness, the quality of different point estimates, the effect of dynamic termination criteria and many other settings were investigated.

ei

[BibTex]

[BibTex]

2004


no image
Statistical Learning with Similarity and Dissimilarity Functions

von Luxburg, U.

pages: 1-166, Technische Universität Berlin, Germany, Technische Universität Berlin, Germany, 2004 (phdthesis)

ei

PDF PostScript [BibTex]

2004


PDF PostScript [BibTex]


no image
Classification and Feature Extraction in Man and Machine

Graf, AAB.

Biologische Kybernetik, University of Tübingen, Germany, 2004, online publication (phdthesis)

ei

[BibTex]

[BibTex]