139x Filetype PDF File size 2.50 MB Source: smartech.gatech.edu
STATISTICAL METHODS WITH APPLICATIONS TO MACHINELEARNINGANDARTIFICIAL INTELLIGENCE AThesis Presented to The Academic Faculty by Yibiao Lu In Partial Fulfillment of the Requirements for the Degree Doctor of Philosophy in the H. Milton Stewart School of Industrial and Systems Engineering Georgia Institute of Technology August 2012 c Copyright 2012 by Yibiao Lu STATISTICAL METHODS WITH APPLICATIONS TO MACHINELEARNINGANDARTIFICIAL INTELLIGENCE Approved by: Professor Xiaoming Huo, Advisor Professor Alex Shapiro H. Milton Stewart School of Industrial H. Milton Stewart School of Industrial and Systems Engineering and Systems Engineering Georgia Institute of Technology Georgia Institute of Technology Professor Shi-Jie Deng Professor Panagiotis Tsiotras H. Milton Stewart School of Industrial The Daniel Guggenheim School of and Systems Engineering Aerospace Engineering Georgia Institute of Technology Georgia Institute of Technology Professor Ming Yuan Date Approved: Apr 26, 2012 H. Milton Stewart School of Industrial and Systems Engineering Georgia Institute of Technology TABLEOFCONTENTS LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii SUMMARY. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xii I THEORETICALRESULTSONHIGH-ORDERLAPLACIAN-BASED REGULARIZATION IN FUNCTION ESTIMATION . . . . . . . 1 1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 1.2.1 Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 1.2.2 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . 5 1.2.3 Choice of the Penalty Parameter λ . . . . . . . . . . . . . . . 6 1.3 Theoretical Properties . . . . . . . . . . . . . . . . . . . . . . . . . . 6 1.3.1 Mathematical Preparation . . . . . . . . . . . . . . . . . . . 7 1.3.2 Bounds of regularization matrix M’s eigenvalues . . . . . . . 11 1.3.3 Convergence Rate of Multivariate GLS Estimator . . . . . . 13 1.3.4 Asymptotic Optimality of GCV . . . . . . . . . . . . . . . . 13 1.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 1.5 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 1.5.1 Detailed Proofs . . . . . . . . . . . . . . . . . . . . . . . . . 17 1.5.2 Agmon’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . 36 1.5.3 Neumann Boundary Condition . . . . . . . . . . . . . . . . . 37 II BEAMLET-BASED GRAPH STRUCTURE FOR PATH PLAN- NINGUSINGMULTISCALEINFORMATION . . . . . . . . . . 38 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 2.2 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . 41 2.3 Multiscale Path Planning Strategy with Preprocessed Information . 43 2.3.1 Recursive Dyadic Partitioning of the Environment . . . . . . 44 2.3.2 Beamlet-like Connectivity . . . . . . . . . . . . . . . . . . . . 46 iii 2.3.3 Bottom-Up Fusion Algorithm . . . . . . . . . . . . . . . . . . 50 2.3.4 Multiscale A* Algorithm on the Beamlet Graph . . . . . . . 52 2.4 Complexity Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 2.4.1 Complexity of Information Fusion Part . . . . . . . . . . . . 53 2.4.2 Complexity of Searching . . . . . . . . . . . . . . . . . . . . 57 2.4.3 Memory Usage . . . . . . . . . . . . . . . . . . . . . . . . . . 59 2.4.4 Preprocessing Time . . . . . . . . . . . . . . . . . . . . . . . 60 2.5 Numerical Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 2.5.1 Comparison Based on L1 Heuristic . . . . . . . . . . . . . . . 61 2.5.2 Comparison Using Stronger Heuristics . . . . . . . . . . . . . 64 2.6 Discussion and Related Prior Work . . . . . . . . . . . . . . . . . . 65 2.6.1 Beamlets as a Predecessor . . . . . . . . . . . . . . . . . . . 65 2.6.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . 66 2.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 III AN INCREMENTAL, MULTI-SCALE SEARCH ALGORITHM FORDYNAMICPATHPLANNINGWITHLOWWORST-CASE COMPLEXITY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 3.2 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . 77 3.3 The Multiscale A* and Lifelong Planning A* Algorithms . . . . . . 79 3.3.1 The Multiscale A* (m-A*) Algorithm . . . . . . . . . . . . . 79 3.3.2 Incremental Search Algorithm: LPA* . . . . . . . . . . . . . 82 3.4 Multiscale Strategy in Dynamic Path Planning: m-LPA* . . . . . . 84 3.4.1 Dynamic Path-Finding Reduced Recursive Dyadic Partition . 84 3.4.2 Update of Multiscale Information in the Beamlet Graph . . . 85 3.4.3 LPA* Algorithm on the Beamlet Graph . . . . . . . . . . . . 87 3.5 Complexity Analysis and Data Structure . . . . . . . . . . . . . . . 90 3.5.1 Worst-Case Complexity Analysis . . . . . . . . . . . . . . . . 90 3.5.2 Fibonacci vs Binomial Heap Implementation . . . . . . . . . 91 iv
no reviews yet
Please Login to review.