Autonomous Learning Systems

Prologue 7 1. Introduction 10 1.1 Autonomous Systems 13 1.2 The Role of Machine Learning in Autonomous Systems 15 1.3 System Identification 20 1.3.3 Novelty detection, outliers and the link to structure innovation 20 1.4 On-line versus Off-line Identification 21 1.5 Adaptive and Evolving Systems 22 1.6 Evolving or Evolutionary Systems 23 1.7 Supervised versus Un-supervised Learning 25 1.8 Structure of the Book 26 PART I: Fundamentals 29 2. Fundamentals of Probability Theory 29 2.1 Randomness and Determinism 30 2.2 Frequentistic versus Belief-based Approach 33 2.3 Probability Densities and Moments 33 2.4 Density Estimation -- Kernel-based Approach 37 2.5 Recursive Density Estimation (RDE) 40 2.6 Detecting Novelties/Anomalies/Outliers using RDE 45 2.7 Conclusion 49 3. Fundamentals of Machine Learning and Pattern Recognition 51 3.1 Pre-processing 51 3.1.1 Normalisation and standardisation 52 3.1.2 Orthogonalization of inputs/features -- rPCA method 53 3.2 Clustering 56 3.2.1 Proximity measures and clusters shape 59 3.2.2 Off-line methods 60 3.2.3 Evolving clustering methods 65 3.3 Classification 73 3.3.1 Recursive LDA, rLDA 74 3.4 Conclusion 74 4. Fundamentals of Fuzzy Systems Theory 77 4.1 Fuzzy Sets 77 4.2 Fuzzy Systems, Fuzzy Rules 80 4.2.1 Fuzzy Systems of Zadeh-Mamdani Type 81 4.2.2 Takagi-Sugeno Fuzzy Systems 83 4.3 Fuzzy Systems with Non-parametric Antecedents (AnYa) 87 4.3.1 Architecture 87 4.3.2 Analysis of AnYa 90 4.4 FRB (Off-line) classifiers 91 4.5 Neuro-Fuzzy Systems 94 4.5.1 Neuro-fuzzy system architecture 94 4.5.2 Evolving NFS as a framework for autonomous learning and knowledge extraction from data streams 97 4.5.3 Linguistic interpretation of the NFS 98 4.6 State Space Perspective 99 4.7 Conclusions 100 Part II Methodology of Autonomous Learning Systems 101 5 Evolving System Structure from Streaming Data 101 5.1 Defining system structure based on prior knowledge 101 5.2 Data Space Partitioning 102 5.2.1 Regular partitioning of the data space 103 5.2.2 Data space partitioning through clustering 104 5.2.3. Data space partitioning based on data clouds 105 5.2.4. Importance of partitioning the joint input-output data space 105 5.2.5 Principles of data space partitioning for autonomous machine learning 107 5.2.6 Dynamic data space partitioning -- evolving system structure autonomously 108 5.3 Normalisation and Standardisation of Streaming Data in Evolving Environments 114 5.3.1 Standardization in an Evolving Environment 115 5.3.2 Normalisation in an Evolving Environment 116 5.4 Autonomous Monitoring of the Structure Quality 117 5.4.1 Autonomous Input Variables Selection 117 5.4.2 Autonomous Monitoring of the Age of the Local Sub-model 120 5.4.3 Autonomous Monitoring of the Utility of the Local Sub-model 122 5.4.4 Update of the Cluster Radii 123 5.5 Short- and Long-term Focal Points and Sub-models 124 5.6 Simplification and Interpretability Issues 125 5.7 Conclusion 127 6 Autonomous Learning Parameters of the Local Sub-models 129 6.1 Learning Parameters of Local Sub-models 130 6.2 Global versus Local Learning 131 6.3 Evolving Systems Structure Recursively 133 6.4 Learning Modes 137 6.5 Robustness to Outliers in Autonomous Learning 140 6.6 Conclusions 140 7 Autonomous Predictors, Estimators, Filters, Inferential Sensors 142 7.1 Predictors, Estimators, Filters -- Problem Formulation 142 7.2 Non-linear Regression 144 7.3 Time series 145 7.4 Autonomous Learning Sensors 146 7.4.1 Autonomous Sensors -- Problem Definition 146 7.4.2 A brief Overview of Soft/Intelligent/Inferential Sensors 147 7.4.3 Autonomous Intelligent Sensors (AutoSense) 149 7.4.4 AutoSense Architecture 151 7.4.5 Modes of Operation of AutoSense 152 7.4.6 Autonomous Input Variable Selection 152 7.5 Conclusions 153 8. Autonomous Learning Classifiers 155 8.1 Classifying data streams 155 8.2 Why adapt the classifier structure? 155 8.3 Architecture of Autonomous Classifiers of the Family AutoClassify 157 8.3.1 AutoClassify0 159 8.3.2 AutoClassify1 159 8.4 Learning AutoClassify from Streaming Data 162 8.4.1 Learning AutoClassify0 162 8.4.2 Learning AutoClassify1 163 8.5 Analysis of AutoClassify methods 163 8.6 Conclusions 164 9. Autonomous Learning Controllers 166 9.1 Indirect Adaptive Control Scheme 167 9.2 Evolving Inverse Plant Model from On-line Streaming Data 169 9.3 Evolving Fuzzy Controller Structure from On-line Streaming Data 170 9.4 Examples of using AutoControl 172 9.5 Conclusion 177 10. Collaborative Autonomous Learning Systems 179 10.1 Distributed Intelligence Scenarios 179 10.2 Autonomous Collaborative Learning 181 10.3 Collaborative Autonomous Clustering, AutoCluster by a team of ALSs 183 10.4 Collaborative Autonomous Predictors, Estimators, Filters and AutoSense by a team of ALSs 184 10.5 Collaborative Autonomous Classifiers AutoClassify by a team of ALSs 184 10.6 Superposition of Local Sub-models 185 10.7 Conclusion 186 PART III: Applications of ALS 187 11. Autonomous Learning Sensors for Chemical and Petro-chemical Industries 187 11.1 Case Study 1: Quality of the Products in an Oil Refinery 187 11.1.1 Introduction 187 11.1.2 The current state of the art 188 11.1.3 Problem description 189 11.1.4 The data set 189 11.1.5 AutoSesnse for kerosene quality prediction 191 11.1.6 AutoSense for Abel inflammability test 193 11.2 Case Study 2: Polypropylene Manufacturing 194 11.2.1 Problem description 194 11.2.2 Drift and shift detection by cluster age derivatives 198 11.2.3 Input variables selection 200 11.3 Conclusion 201 12. Autonomous Learning Systems in Mobile Robotics 203 12.1 The mobile robot Pioneer 3DX 203 12.2 Autonomous Classifier for Landmark Recognition 205 12.2.1 Corner detection and simple mapping of an indoor environment through wall following 207 12.2.2 Outdoor landmark detection based on visual input information 210 12.2.3 VideoDiaries 214 12.2.4 Collaborative scenario 217 12.3 Autonomous Leader Follower 220 12.4 Results Analysis 223 13. Autonomous Novelty Detection and Object Tracking in Video Streams 224 13.1 Problem Definition 224 13.2 Background subtraction and KDE for detecting visual novelties 225 13.2.1 Background subtraction method 225 13.2.2 Challenges 226 13.2.3 Parametric versus non-parametric approaches 229 13.2.4 Kernel Density Estimation method 230 13.3 Detecting Visual novelties with RDE Method 231 13.4 Object Identification in Image Frames using RDE 232 13.5 Real-time Tracking in Video Streams using ALS 234 13.6 Conclusion 237 14. Modelling Evolving User Behaviour with ALS 239 14.1 User Behaviour as an evolving phenomenon 239 14.2 Designing the User Behaviour Profile 241 14.3 Applying AutoClassify0 for modelling evolving user behaviour 244 14.4 Case studies 245 14.4.1 Users of UNIX commands 245 14.4.2 Modelling activity of people in a smart home environment 247 14.4.3 Automatic scene recognition 249 14.5 Conclusions 252 15. Epilogue 254 15.1 Conclusions 254 15.2 Open Problems 258 15.3 Future Directions 259 Bibliography 261 Index 274 Glossary 291 Appendices 295 A. Mathematical Foundations 296 A1 Probability distributions 297 A2 Basic matrix properties 300 B. Pseudo-code of the basic algorithms 302 B1 Mean shift with Epanechnikov kernel 302 B2 AutoCluster 304 B3 ELM 305 B4 AutoPredict 307 B5 AutoSense 308 B6 AutoClassify0 309 B7 AutoClassify1 311 B8 AutoControl 313