[TMLR 2026] A Multi-Fidelity Control Variate Approach forPolicy Gradient Estimation скачать в хорошем качестве

[TMLR 2026] A Multi-Fidelity Control Variate Approach forPolicy Gradient Estimation

Presentation of our TMLR 2026 paper on Multi-Fidelity Policy Gradient (MFPG) TL;DR: MFPG is a multi-fidelity, sample-efficient reinforcement learning (RL) framework that boosts RL with cheap, imperfect data. Key properties: Unbiasedness: grounds learning on scarce, accurate, high-fidelity data Reduced variance: uses large amounts of cheap, imperfect, low-fidelity data as a control variate for variance reduction Robustness: robust to low-fidelity data biases Generality: handles dynamics gaps and reward misspecification For more information, please consult our website: https://xinjie-liu.github.io/mfpg-rl/ We welcome discussions, feedback, and questions!