Skip to yearly menu bar Skip to main content


( events)   Timezone:  
Poster
Wed May 08 09:00 AM -- 11:00 AM (PDT) @ Great Hall BC #36
Information asymmetry in KL-regularized RL
Alexandre Galashov · Siddhant Jayakumar · Leonard Hasenclever · Dhruva Tirumala · Jonathan Schwarz · Guillaume Desjardins · Wojciech M Czarnecki · Yee Whye Teh · Razvan Pascanu · Nicolas Heess
[ PDF

Many real world tasks exhibit rich structure that is repeated across different parts of the state space or in time. In this work we study the possibility of leveraging such repeated structure to speed up and regularize learning. We start from the KL regularized expected reward objective which introduces an additional component, a default policy. Instead of relying on a fixed default policy, we learn it from data. But crucially, we restrict the amount of information the default policy receives, forcing it to learn reusable behaviors that help the policy learn faster. We formalize this strategy and discuss connections to information bottleneck approaches and to the variational EM algorithm. We present empirical results in both discrete and continuous action domains and demonstrate that, for certain tasks, learning a default policy alongside the policy can significantly speed up and improve learning. Please watch the video demonstrating learned experts and default policies on several continuous control tasks ( https://youtu.be/U2qA3llzus8 ).