Inference and Distillation for Option Learning

Maximilian Igl, Wendelin Boehmer, Andrew Gambardella, Philip H.S. Torr, Nantas Nardelli, N. Siddharth, Shimon Whiteson

Research output: Contribution to conferencePaperpeer-review

Abstract

We present Inference and Distillation for Option Learning (IDOL), a multitask option-learning framework based on Planning-as-Inference. IDOL employs a hierarchical prior and variational-posterior factorisation to learn temporally extended options that allow the higher-level master policy to make decisions with lower frequency, speeding up training on new tasks. IDOL autonomously learns the temporal extension of each option and avoids suboptimal solutions where multiple options learn similar behavior. We demonstrate that this improves performance on new tasks compared to both strong hierarchical and flat transfer-learning baselines.
Original languageEnglish
Number of pages10
Publication statusPublished - 8 Dec 2018
EventWorkshop on Probabilistic Reinforcement Learning and Structured Control @ NeurIPS 2018: Infer to Control - Montréal, Canada
Duration: 8 Dec 20188 Dec 2018
https://sites.google.com/view/infer2control-nips2018/home?authuser=0

Workshop

WorkshopWorkshop on Probabilistic Reinforcement Learning and Structured Control @ NeurIPS 2018
Abbreviated titleInfer2Control @ NeurIPS 2018
Country/TerritoryCanada
CityMontréal
Period8/12/188/12/18
Internet address

Fingerprint

Dive into the research topics of 'Inference and Distillation for Option Learning'. Together they form a unique fingerprint.

Cite this