Reinforcement Learning formalises an embodied agent's interaction with the environment through observations, rewards and actions. But where do the actions come from? Actions are often considered to represent something external, such as the movement of a limb, a chess piece, or more generally, the output of an actuator. In this work we explore and formalize a contrasting view, namely that actions are best thought of as the output of a sequence of internal choices with respect to an action model. This view is particularly well-suited for leveraging the recent advances in large sequence models as prior knowledge for multi-task reinforcement learning problems. Our main contribution in this work is to show how to augment the standard MDP formalism with a sequential notion of internal action using information-theoretic techniques, and that this leads to self-consistent definitions of both internal and external action value functions.

Hierarchical reinforcement learning (HRL) helps address large-scale and sparse reward issues in reinforcement learning. In HRL, the policy model has an inner representation structured in levels. With this structure, the reinforcement learning task is expected to be decomposed into corresponding levels with sub-tasks, and thus the learning can be more efficient. In HRL, although it is intuitive that a high-level policy only needs to make macro decisions in a low frequency, the exact frequency is hard to be simply determined. Previous HRL approaches often employed a fixed-time skip strategy or learn a terminal condition without taking account of the context, which, however, not only requires manual adjustments but also sacrifices some decision granularity. In this paper, we propose the \emph{temporal-adaptive hierarchical policy learning} (TEMPLE) structure, which uses a temporal gate to adaptively control the high-level policy decision frequency. We train the TEMPLE structure with PPO and test its performance in a range of environments including 2-D rooms, Mujoco tasks, and Atari games. The results show that the TEMPLE structure can lead to improved performance in these environments with a sequential adaptive high-level control.