| Abstract [eng] |
Irregularly-sampled time series occur in many real-world applications, such as robots movement, astronomy, medicine. In this master's thesis, we concentrate on learning robot movement trajectories with imitation learning type of algorithms. Recently, a huge interest has been in the methods that perform end-to-end trajectory learning directly using irregularly-sampled time series as input without any additional data preparation steps because data gaps are informative themself. Such methods are usually based on Recurrent Neural Networks (RNNs). In this thesis, we propose the new deep learning model called Ordinary Differential Equations based Gated Recurrent Unit with Trainable Decays (ODE-GRU-D), which is based on state-of-the-art RNN models ODE-RNN and Gated Recurrent Unit with trainable Decays (GRU-D). Additionally, three other commonly used algorithms are selected for the experiments: one standard RNN model - Gated Recurrent Unit (GRU), and two specific algorithms for irregularly-sampled time series - GRU-D and ODE-RNN. The models are applied and compared on the irregularly-sampled MuJoCo Hopper trajectories datasets. ODE-GRU-D showed several advantages: it is more stable and converges faster than other investigated RNNs. Both ODE-based models notably outperformed the other two. Also, the proposed algorithm achieved significantly better results than the previous ODE-based model ODE-RNN on sparse time series cases. |
