Human Behavior Learning and Transfer

Bridging The Gap Between Human-computer Engineering And Control Engineering. Human Behavior Learning And Transfer Delineates How To Abstract Human Action And Reaction Skills Into Computational Models. The Authors Include Methods For Modeling A Variety Of Human Action And Reaction Behaviors And Explore Processes For Evaluating, Optimizing, And Transferring Human Skills. They Also Cover Modeling Continuous And Discontinuous Human Control Strategy And Discuss Simulation Studies And Practical Real-life Situations.--jacket. Introduction To Human Reaction Skill Modeling -- Learning Of Human Control Strategy : Continuous And Discontinuous -- Validation Of Human Control Strategy Models -- Evaluation Of Human Control Strategy -- Performance Optimization Of Human Control Strategy -- Transfer Of Human Control Strategy -- Transferring Human Navigational Skills To Smart Wheelchair -- Introduction To Human Action Skill Modeling -- Global Parametric Methods For Dimension Reduction -- Local Methods For Dimension Reduction -- A Spline Smoother In Phase Space For Trajectory Fitting -- Analysis Of Human Walking Trajectories For Surveillance -- Modeling Of Facial And Full-body Actions -- Appendix A : Human Control Data. Yangsheng Xu, Ka Keung Lee. Includes Bibliographical References (p. 309-329) And Index. FM Human Behavior Learning and Transfer 2 Preface 5 Authors 7 Contents 9 Chapter 1: Introduction 14 1.1 Motivation 14 1.2 Overview 15 Chapter 2: Introduction to Human Reaction Skill Modeling 19 2.1 Motivation 19 2.2 Related work 21 2.2.1 Skill learning through exploration 21 2.2.2 Skill modeling from human data 22 2.2.3 Neural network learning 23 2.2.4 Locally weighted learning 25 Chapter 3: Learning of Human Control Strategy: Continuous and Discontinuous 27 3.1 Experimental design 27 3.1.1 Motivation 27 3.1.2 Simulation environment 28 Dynamic driving simulator 28 Road descriptions 30 3.1.3 Model class 30 3.2 Cascade neural networks with Kalman filtering 33 3.2.1 Cascade neural networks 33 3.2.2 Node-decoupled extended Kalman filtering 35 Learning architecture 35 3.3 HCS models: continuous control 37 3.3.1 Cascade with quickprop learning 37 Experimental data 37 Model inputs and outputs 37 Cq training 38 HCS models 38 3.3.2 Cascade with NDEKF learning 43 Model inputs and outputs 43 Ck training 43 HCS models 44 3.3.3 Analysis 49 Model stability 49 Learning convergence 49 Discussion 53 3.4 HCS models: discontinuous control 55 3.4.1 Hybrid continuous/discontinuous control 55 General statistical framework 55 Action definitions 57 Statistical model choice 58 Prior probabilities 59 Task-based modifications 59 3.4.2 Experimental results 61 Model training 61 HCS models 61 3.4.3 Analysis 66 Sample curve control 66 Probability profile 69 Modeling extension 70 Chapter 4: Validation of Human Control Strategy Models 74 4.1 Need for model validation 74 4.2 Stochastic similarity measure 77 4.2.1 Hidden Markov models 78 4.2.2 Similarity measure 79 4.2.3 Properties 81 4.2.4 Distance measure 83 4.2.5 Data preprocessing 84 4.2.6 Vector quantization 89 4.2.7 Discretization compensation 92 4.2.8 HMM training 95 4.3 Human-to-model comparisons 96 Chapter 5: Evaluation of Human Control Strategy 100 5.1 Introduction 100 5.2 Obstacle avoidance 101 5.2.1 Virtual path equivalence 102 5.2.2 Lateral offset estimation 103 5.2.3 Obstacle avoidance threshold 104 5.2.4 Obstacle avoidance velocity loss 106 5.3 Tight turning 107 5.3.1 Tight turning connections 108 5.3.2 Threshold with tight angle 109 5.4 Transient response 111 5.5 Time delay 114 5.6 Passenger comfort 116 5.7 Driving smoothness 122 5.8 Summary 124 Chapter 6: Performance Optimization of Human Control Strategy 125 6.1 Introduction 125 6.2 Simultaneously perturbed stochastic approximation 126 6.3 Iterative optimization algorithm 128 6.4 Model optimization and performance analysis 131 6.5 Summary 134 Chapter 7: Transfer of Human Control Strategy 135 7.1 Introduction 135 7.2 Model transfer based on similarity measure 136 7.2.1 Structure learning 137 7.2.2 Parameter learning 137 7.2.3 Experimental study 139 7.3 Model compensation 142 7.4 Summary 146 Chapter 8: Transferring Human Navigational Skills to Smart Wheelchair 147 8.1 Introduction 147 8.1.1 Related work 147 8.2 Methodology 149 8.2.1 Problem formulation 149 8.2.2 Theoretical foundation 150 8.3 Experimental study 151 8.3.1 Settings 151 8.3.2 Experiment 1: Navigation 152 Procedure 152 Results 155 8.3.3 Experiment 2: Localization 156 Procedure 156 Result 1: Localization performance 160 Result 2: Similar sensor patterns in various configurations 160 Result 3: Small variations of major dimensions of environmental features along the route 160 8.4 Analysis 163 8.4.1 Performance evaluation 163 8.4.2 Advantages of the approach 164 8.4.3 Choices of sensor-configuration mapping 165 8.4.4 Generalization of the study 166 8.5 Conclusion 166 Chapter 9: Introduction to Human Action Skill Modeling 168 9.1 Learning action models from human demonstrations 168 9.2 Formulation of the dimension reduction problem 171 9.3 Related research 174 Primitives and skills 174 Dimension reduction and local learning 175 Uses for action models 176 Chapter 10: Global Parametric Methods for Dimension Reduction 178 10.1 Introduction 178 10.2 Parametric methods for global modeling 179 10.2.1 Polynomial regression 179 10.2.2 First principal component 180 10.3 An experimental data set 182 10.4 Principal component analysis for modeling human performance data 184 10.5 NLPCA 186 10.6 SNLPCA 192 10.7 Comparison 195 10.8 Characterizing NLPCA mappings 196 Chapter 11: Local Methods for Dimension Reduction 201 11.1 Introduction 201 11.2 Local, non-parametric methods for trajectory fitting 202 11.3 Scatter plot smoothing 203 11.4 Action recognition using smoothing splines 205 11.5 A gesture-recognition experiment using spline smoothing 207 11.6 Principal curves 210 11.6.1 Definition of principal curves 212 11.6.2 Distance property 212 11.6.3 Principal curves algorithm for distributions 213 11.6.4 Principal curves algorithm for data sets: projection step 214 11.6.5 Principal curves algorithm for datasets: conditional expectation step 216 11.7 Expanding the one-dimensional representation 217 11.8 Branching 219 11.9 Over-fitting 221 Chapter 12: A Spline Smoother in Phase Space for Trajectory Fitting -1 12.1 Trajectory smoothing with velocity information 222 12.2 Problem formulation 224 12.3 Solution 226 12.4 Notes on computation and complexity 230 12.5 Combining points with similar parameterizations 232 12.6 Multi-dimensional smoothing 233 12.7 Estimation of variances 234 12.8 Windowing variance estimates 236 12.9 The effect of velocity information 237 12.10 Cross-validation 237 Chapter 13: Analysis of Human Walking Trajectories for Surveillance 240 13.1 Introduction 240 13.2 System overview 242 13.3 Background subtraction 242 13.4 Local trajectory point classification 243 13.5 Global trajectory similarity estimation 245 13.6 Trajectory normality classifier 247 13.7 Experiment 1: Trajectory normality classifier 247 13.8 Further analysis on global trajectory similarity based on LCSS 251 13.9 Methodology used in boundary modeling 251 13.9.1 Trajectory similarity based on LCSS 252 13.10 LCSS boundary limit establishment 253 13.10.1 LCSS thresholds learnt from support vector regression 254 13.10.2 LCSS thresholds learnt from cascade neural networks 254 13.10.3 Fixed LCSS thresholds 255 13.10.4 Variable LCSS thresholds 255 13.11 Experiment 2: Boundary modeling 256 13.12 Discussion 265 13.13 Conclusion 266 Chapter 14: Modeling of Facial and Full-Body Actions 267 14.1 Facial expression intensity modeling 267 14.1.1 Related work 269 14.1.2 System overview 271 14.1.3 Extraction of facial motion data 271 14.1.4 Automatic extraction of facial expression intensity from transition 273 1. Representation of manifold 274 2. Computation of the shortest distances in the manifold 274 3. Construction of embedding 274 14.1.5 The learning of facial expression intensity 275 Cascade neural network 275 Support vector machine 276 14.1.6 Experiment 278 14.1.7 Discussion 280 14.1.8 Conclusion 282 14.2 Full-body action modeling 283 14.2.1 System overview 284 Data preprocessing 284 14.2.2 Feature extraction 286 14.2.3 Learning system based on support vector classification 288 14.2.4 Experiment 1: Recognition of motions of table tennis players 288 14.2.5 Experiment 2: Detection of fighting actions 291 14.2.6 Discussion 294 14.2.7 Conclusion 296 Chapter 15: Conclusions 297 Appendix A: Human Control Data -1 A.1 Larry 303 A.2 Curly 305 A.3 Moe 307 A.4 Groucho 309 A.5 Harpo 311 A.6 Zeppo 313 References 315 Bridging the gap between human-computer engineering and control engineering, the book delineates how to abstract human action and reaction skills into computational models. The authors include methods for modeling a variety of human action and reaction behaviors and explore processes for evaluating, optimizing, and transferring human skills. They also cover modeling continuous and discontinuous human control strategy and discuss simulation studies and practical real-life situations.--ERGONOMICSnetBASE description The authors (both of the Chinese U. of Hong Kong) examine the modeling and transfer of human action and reaction behaviors. They apply machine learning techniques and statistical analysis towards abstracting models of human reaction behavior, contending that such models can be learnt efficiently to emulate complex human control behaviors in the fee Delineates how to abstract human action and reaction skills into computational models. This book includes methods for modeling human action and reaction behaviors and explores processes for evaluating, optimizing, and transferring human skills.