This dissertation is about the application of machine learning to robot control. A system which has no initial model of the robot/world dynamics should be able to construct such a model using data received through its sensors--an approach which is formalized here as the $AB (State-Action-Behaviour) control cycle. A method of learning is presented in which all the experiences in the lifetime of the robot are explicitly remembered. The experiences are stored in a manner which permits fast recall of the closest previous experience to any new situation, thus permitting very quick predictions of the effects of proposed actions and, given a goal behaviour, permitting fast generation of a candidate action. The learning can take place in high-dimensional non-linear control spaces with real-valued ranges of variables. Furthermore, the method avoids a number of shortcomings of earlier learning methods in which the controller can become trapped in inadequate performance which does not improve. Also considered is how the system is made resistant to noisy inputs and how it adapts to environmental changes. A well founded mechanism for choosing actions is introduced which solves the experiment/perform dilemma for this domain with adequate computational efficiency, and with fast convergence to the goal behaviour. The dissertation explefins in detail how the $AB control cycle can be integrated into both low and high complexity tasks. The methods and algorithms are evaluated with numerous experiments using both real and simulated robot domefins. The final experiment also illustrates how a compound learning task can be structured into a hierarchy of simple learning tasks.

In order to support the maintenance of an objectoriented software system, the quality of its design must be evaluated using adequate quantification means. In spite of the current extensive use of metrics, if used in isolation metrics are oftentimes too fine grained to quantify comprehensively an investigated design aspect (e.g., distribution of system’s intelligence among classes). To help developers and maintainers detect and localize design problems in a system, we propose a novel mechanism – called detection strategy – for formulating metrics-based rules that capture deviations from good design principles and heuristics. Using detection strategies an engineer can directly localize classes or methods affected by a particular design flaw (e.g., God Class), rather than having to infer the real design problem from a large set of abnormal metric values. We have defined such detection strategies for capturing around ten important flaws of object-oriented design found in the literature and validated the approach experimentally on multiple large-scale case-studies.

The psychometric function relates an observer’s performance to an independent variable, usually some physical quantity of a stimulus in a psychophysical task. This paper, together with its companion paper (Wichmann & Hill, 2001), describes an integrated approach to (1) fitting psychometric functions, (2) assessing the goodness of fit, and (3) providing confidence intervals for the function’s parameters and other estimates derived from them, for the purposes of hypothesis testing. The present paper deals with the first two topics, describing a constrained maximum-likelihood method of parameter estimation and developing several goodness-of-fit tests. Using Monte Carlo simulations, we deal with two specific difficulties that arise when fitting functions to psychophysical data. First, we note that human observers are prone to stimulus-independent errors (or lapses). We show that failure to account for this can lead to serious biases in estimates of the psychometric function’s parameters and illustrate how the problem may be overcome. Second, we note that psychophysical data sets are usually rather small by the standards required by most of the commonly applied statistical tests. We demonstrate the potential errors of applying traditional c 2 methods to psychophysical data and advocate use of Monte Carlo resampling techniques that do not rely on asymptotic theory. We have made available the software to implement our methods. The performance of an observer on a psychophysical

A fundamental problem in depth from defocus is the measurement of relative defocus between images. The performance of previously proposed focus operators are inevitably sensitive to the frequency spectra of local scene textures. As a result, focus operators such as the Laplacian of Gaussian result in poor depth estimates. An alternative is to use large filter banks that densely sample the frequency space. Though this approach can result in better depth accuracy, it sacrifices the computational efficiency that depth from defocus offers over stereo and structure from motion. We propose a class of broadband operators that, when used together, provide invariance to scene texture and produce accurate and dense depth maps. Since the operators are broadband, a small number of them are sufficient for depth estimation of scenes with complex textural properties. In addition, a depth confidence measure is derived that can be computed from the outputs of the operators. This confidence measure perm...

We propose an approach (GQM/MEDEA) for defining measures of product attributes in software engineering. The approach is driven by the experimental goals of measurement, expressed via the GQM paradigm, and a set of empirical hypotheses. To make

This article gives a concise overview of Bayesian sampling for neural networks, and then presents an extensive evaluation on a set of various benchmark classification problems. The main objective is to study the sensitivity of this scheme to changes in the prior distribution of the parameters and hyperparameters, and to evaluate the efficiency of the so-called automatic relevance determination (ARD) method. The paper concludes with a comparison of the achieved classification results with those obtained with (i) the evidence scheme and (ii) with non-Bayesian methods. Keywords Bayesian statistics, prior and posterior distribution, parameters and hyperparameters, Gibbs sampling, hybrid Monte Carlo, automatic relevance determination (ARD), evidence approximation, classification problems, benchmarking. 1 Theory: Sampling of network weights and hyperparameters from the posterior distribution The objective of this section is to give a concise yet self-contained overview of the Bayesian app...

We explain how to apply statistical techniques to solve several language-recognition problems that arise in cryptanalysis and other domains. Language recognition is important in cryptanalysis because, among other applications, an exhaustive key search of any cryptosystem from ciphertext alone requires a test that recognizes valid plaintext. Written for cryptanalysts, this guide should also be helpful to others as an introduction to statistical inference on Markov chains. Modeling language as a finite stationary Markov process, we adapt a statistical model of pattern recognition to language recognition. Within this framework we consider four welldefined language-recognition problems: 1) recognizing a known language, 2) distinguishing a known language from uniform noise, 3) distinguishing unknown 0th-order noise from unknown 1st-order language, and 4) detecting non-uniform unknown language. For the second problem we give a most powerful test based on the Neyman-Pearson Lemma. For the oth...

In this paper we define two alternatives to the familiar perplexity statistic (hereafter lexical perplexity), which is widely applied both as a measure-of-goodness and as an objective function for training language models. These alternatives, respectively acoustic perplexity and the synthetic acoustic word error rate, fuse information from both the language model and the acoustic model. We show how to compute these statistics by effectively synthesizing a large acoustic corpus, demonstrate their superiority to lexical perplexity as predictors of language model performance, and investigate their use as objective functions for training language models. We present results from a simple speech recognition experiment that demonstrate a small reduction in word error rate.

For conventional single-event, nonlinear, least-squares hypocentral estimates, I show that the total error is expressible as a linear combination of three terms: (1) measurement error; (2) modeling errors caused by inadequacy of the travel-time tables; and (3) a nonlinear term. Errors in calculating travel-time partial derivatives are shown to have no effect, provided a stable solution can be found. This is in contrast to linear problems where errors in calculating matrix elements can distort the solution drastically. The error appraisal technique developed here examines each of the three error terms independently. The first can be analyzed by standard confidence ellipsoids with critical values based on measurement error statistics. The second can cause conventional error ellipsoid calculations that derive a critical value from an estimate based on rms residuals, to give misleading results. I introduce an alternative extremal bound procedure for appraising such errors. Travel-time modeling errors are bounded as the product of ray arc length and an estimate of the nominal scale of slowness errors along the ray path. These are used to derive an upper bound on systematic errors in each hypocentral coordinate based on a novel bounding criteria. Finally, I show that, for errors of a reasonable scale, the nonlinear error term can be estimated adequately using a second-order approximation. Given an upper bound on the total location error, bounds on the travel-time error induced by nonlinearity can be calculated from the spectral norm of the Hessian for each measured arrival time. The systematic errors in each hypocentral coordinate due to nonlinearity can then be bounded using the same criteria used for constructing modeling error bounds. This overall procedure is complete because it allows one to independently appraise the relative importance of all sources of hypocentral errors. It is practical because the required computational effort is small.

xiii Declaration, xiv The Author, xv Acknowledgements, xv 1 Chapter 1 -