This website requires JavaScript.

SHIRO: Soft Hierarchical Reinforcement Learning

Kandai WatanabeMathew StrongOmer Eldar
Dec 2022
Hierarchical Reinforcement Learning (HRL) algorithms have been demonstratedto perform well on high-dimensional decision making and robotic control tasks.However, because they solely optimize for rewards, the agent tends to searchthe same space redundantly. This problem reduces the speed of learning andachieved reward. In this work, we present an Off-Policy HRL algorithm thatmaximizes entropy for efficient exploration. The algorithm learns a temporallyabstracted low-level policy and is able to explore broadly through the additionof entropy to the high-level. The novelty of this work is the theoreticalmotivation of adding entropy to the RL objective in the HRL setting. Weempirically show that the entropy can be added to both levels if theKullback-Leibler (KL) divergence between consecutive updates of the low-levelpolicy is sufficiently small. We performed an ablative study to analyze theeffects of entropy on hierarchy, in which adding entropy to high-level emergedas the most desirable configuration. Furthermore, a higher temperature in thelow-level leads to Q-value overestimation and increases the stochasticity ofthe environment that the high-level operates on, making learning morechallenging. Our method, SHIRO, surpasses state-of-the-art performance on arange of simulated robotic control benchmark tasks and requires minimal tuning.