Comparison of Independently Learnable Hierarchical Models for Time Series Data

Abstract

In the hierarchical imitation learning model, multiple neural networks (NN) with memory can learn long horizon tasks. However, because the upper and lower layers have memories, independently learnable models may have an inconsistency between the phases of operation predicted by the upper and lower layer models, which may have a negative effect on motion. Since the lower layers are given the current observation and the future predictions, we thought the motion generation is predictable without considering past state. In this paper, we examined the effect on motion and learning time when the lower layer have no memory.