AI Paper Atlas

Research map / Reinforcement learning / Policy gradient and actor-critic

Expected Policy Gradients for Reinforcement Learning

Kamil Ciosek, Shimon Whiteson

JMLR 2020 · Reinforcement learning › Policy gradient and actor-critic · approach: Variance-reduced policy gradients · 65 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use EPG covariance-based exploration instead of Ornstein-Uhlenbeck noise in DPGDPG · MuJoCo continuous control (5 tasks)learning curve returnImproved
replace exponential of Hessian with 1-step approximation max(0,1+lambda) (ablation)EPG exploration · MuJoCo continuous control (5 tasks)learning curve returnNo clear effect
use full Hessian instead of diagonal Hessian (ablation)EPG · HalfCheetah, Reacher, Walker2dlearning curve returnNo clear effect
use state-dependent covariance instead of one global averaged Hessian (ablation)EPG · MuJoCo continuous control (5 tasks)learning curve returnImproved
use numerical quadrature EPG actor update instead of DPG/SPG updateDPG with OU noise 0.2 · MuJoCo continuous control (5 tasks)learning curve returnMixed
increase quadrature samples from 1 to 4 and 8 (ablation)NQ-EPG · MuJoCo continuous control (5 tasks)learning curve returnNo clear effect
use EPG sum over actions instead of A2C stochastic policy gradientA2C · Atari Ponglearning curve scoreNo clear effect
compare EPG with PPOEPG · HalfCheetah, Walker2dlearning curve return and stabilityMixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy