Improving machine learning methods for solving non-stationary conditions based on data availability, time urgency, and types of change

Abstract

Supervised learning algorithms do not work well when the deployment condition is dissimilar to the training condition. Such non-stationary conditions include covariate shifts and concept shifts. Importance weighted learning (IWL) is used to handle a one-time covariate shift but not frequent shifts and concept shifts. While forgetting addresses concept shifts, it is wasteful in discarding previously learned models. To address these shortfalls, this thesis proposes looking into the three stages of supervised learning and devised pre-learning methods, which deal with data and feature selection; in-learning methods, which modify the learning process; and post-learning methods, which modify the prediction process to compensate for shifts in conditions.The first in-learning method is a transfer learning-based technique that utilizes a limited amount of test data to train further the prediction model pre-trained on general training data. This technique boosted the accuracy of a vocal emotion recognizer by 10%.For applications that require a timely response, we employed a post-learning strategy in the form of local learning. It handles multiple covariate shifts and improves the prediction accuracy in one vocal emotion recognition instance from 88.8% to 93.2%. Local learning also allows the use of feature augmentation to convert a more difficult concept-shift problem into an easier covariate-shift problem. The resulting controller outperforms PID controllers in water shooting control.When data are abundant, we leverage pre-learning methods such as conditionspecific learning, to avoid non-stationary conditions altogether. Using this technique, we developed a semi-automatic snore labeling software that produces good accuracy (0.93 F1-score) and cuts labeling time from hours to minutes. Alternatively, we use deep learning methods to learn features that are robust to shifts.In our ablation study, we showed that features extracted from very deep networks and recurrent networks results in a more accurate and robust snore classification.xv xvi With the advance of computer simulation, unlimited artificial data can be generated to better approximate and cover possible test conditions. We tested this idea in teaching a double-hull welding robot to climb down safely from a high wall through reinforcement learning and achieved a 90% success rate.Finally, from these applications, we distilled a method selection guideline based on data availability, time urgency, and type of shift.