Data-Driven Decision-Making Based on Noisy Data Samples: Studies in the Machine Learning Applications

Modern machine learning research falls into the field of data-driven decision-making, which inevitably requires consideration of the negative effect of noisy samples. However, majority of the existing study normally assumes the collected datasets to be accurate and clean. Consequently, this unrealistic noise-free assumption puts a heavy burden, e.g., data cleansing, in the practical applications of machine learning. To release this burden and make the machine learning approaches more methodologically robust to sample noise, the investigation on the topic of machine learning under noisy datasets is both meaningful and necessary. In this dissertation, we\footnote Throughout the rest of this thesis, the narrator will be referred as "we'' instead of "I''. This is because the dissertation is consisted of research articles, which are conducted in a collaborative setting. Please note that the writing outsides of the articles themselves is my own. offer a systematic way to investigate the problem of learning from noisy datasets. In particular, we first study how to quantify the data sample quality given there are surprisingly few literature that cover this quantification task. We then study the situation of one-shot decision-making based on noisy data samples. After that, the situation of sequential decision-making is investigated. Finally, we study one real application of machine learning, i.e. , reinforcement learning (RL) for the indoor drone flight control, with a system design for accurate data acquisition. For the purposes of reference, we also provide some future research directions in the end.