PhD Research Seminar: Time series and automated machine learning, Truncated Quantile Critics and overestimation bias
First talk: Time Series Forecasting Based on Automated Machine Learning
Speaker: Konstantin Danilov, second-year PhD student, School of Business Informatics
Nowadays, forecasting time series is important in solving a wide range of problems in various spheres of human activity. Researchers resorted to different approaches to achieve the required accuracy of forecasting models, including feature engineering. This presentation will be about a feature engineering method for time series data based on Bayesian optimization. The proposed approach has been experimentally shown to be higly efficient. Some difficulties related to this approach will be discussed.
Second talk: Controlling Overestimation Bias with Truncated Mixture of Continuous Distributional Quantile Critics
Speaker: Alexander Grishin, second-year PhD student, Faculty of Computer Science
According to previous studies, one of the major impediments to accurate off-policy learning is the overestimation bias. During the talk, I will explore our novel way to alleviate the overestimation bias in a continuous control setting. Our method, Truncated Quantile Critics (TQC), blends three ideas: distributional representation of a critic, truncation of critics prediction, and ensembling of multiple critics. We show that all components are key for the achieved performance. Distributional representation combined with truncation allows for arbitrary granular overestimation control, and ensembling further improves the results of our method. TQC significantly outperforms the current state of the art on all environments from the continuous control benchmark suite, demonstrating 25% improvement on the most challenging Humanoid environment.