AI Paper Atlas

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

Multi-Step Generalized Policy Improvement by Leveraging Approximate Models

Lucas Nunes Alegre, Ana L. C. Bazzan, Ann Nowe, Bruno Castro da Silva

NeurIPS 2023 · Reinforcement learning › Policy gradient and actor-critic · approach: Model-based RL with guarantees · 2 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
increase planning horizon h in h-GPI (1-GPI up to 10-GPI) over standard GPI (h=0)h-GPI with SFs · stochastic FourRoom test tasksnormalized return (median, IQM, mean)Improved
add GPI bootstrapping at step h (h-GPI vs SF-MPC without bootstrapping) (ablation)SF-based MPC · stochastic FourRoomnormalized returnImproved
add Gaussian noise to learned SFs (std 0.05, 0.1, 0.2) on standard GPI (h=0) (ablation)GPI · FourRoom with noisy SFsmean normalized returnHurt
increase h to compensate for noise in SFs (ablation)h-GPI · FourRoom with noisy SFsmean normalized returnImproved
use h-GPI instead of GPI over training tasksUSFA-based GPI · Reachernormalized returnImproved
add constraints on value estimates (CGPI) over GPI-STGPI-ST · Reacher and FetchPushnormalized returnNo clear effect
use h-GPI instead of SF-MPC with up to 15 planning stepsSF-MPC · FetchPushnormalized returnImproved
increase h to intermediate values in h-GPIh-GPI · FetchPushnormalized returnMixed

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