[3D4-OS-4b] 自律・創発・汎用AIアーキテクチャ(2)

[3D4-OS-4b-01] Introducing a Call Stack into the RGoal Hierarchical Reinforcement Learning Architecture

〇Yuuji Ichisugi¹, Naoto Takahashi¹, Hidemoto Nakada¹, Takashi Sano² (1. National Institute of Advanced Industrial Science and Technology (AIST), 2. Department of Computer and Information Science, Faculty of Science and Technology, Seikei University)

Keywords:Hierarchical reinforcement learning, Model-based reinforcement learning, Zero-shot learning

Humans can set suitable subgoals in order to achieve some purposes, and furthermore, can set sub-subgoals recursively if needed.
It seems that the depth of the recursion is unlimited.
Inspired by this behavior, we had designed a hierarchical reinforcement learning architecture, the RGoal architecture.
In this paper, we introduce a call stack into the RGoal architecture to increase reusability of subgoals.
We evaluate its performance using a maze with multi-task setting.
The result shows that the convergence speed improves as the maximum stack size increases.