Original title: Curriculum Learning and Imitation Learning for Model-free Control on Financial Time-series
Authors: Woosung Koh, Insu Choi, Yuntae Jang, Gimin Kang, Woo Chang Kim
In their article, researchers delve into the underexplored realm of applying curriculum learning and imitation learning to control tasks with highly unpredictable time-series data in robotics. They experiment both theoretically and empirically, using data augmentation to introduce curriculum learning and policy distillation from an oracle for imitation learning in a complex control task. Their study underscores the promise of curriculum learning as a novel approach for enhancing control-task performance in such scenarios, supported by extensive empirical evidence and analysis. Encouragingly, fine-tuning hyperparameters even favoring the baseline strengthens the case for curriculum learning. However, they caution against indiscriminate use of imitation learning, highlighting its limitations in this context. Overall, the research emphasizes the potential of curriculum learning while urging careful consideration in implementing imitation learning for control tasks involving stochastic time-series data.
Original article: https://arxiv.org/abs/2311.13326