Michał Bortkiewicz

@m_bortkiewicz

PhD at Warsaw University of Technology Working on RL and Continual Learning

michalbortkiewicz.github.io

Joined February 2021

544 Following

474 Followers

Replies Reposts Media

Pinned
Michał Bortkiewicz
@m_bortkiewicz
Sep 12, 2024
I am excited to share our recent work with @WladekPalucki , @vivek_myers, @Taddziarm , @tomArczewski, @LukeKucinski, and @ben_eysenbach! Accelerating Goal-Conditioned Reinforcement Learning Algorithms and Research Webpage: michalbortkiewicz.github.io/JaxGCRL/
1
Michał Bortkiewicz
@m_bortkiewicz
Sep 20
humbling to watch human intellectual supremacy crumble. we grew up with the great mathematician as the pinnacle of reasoning. the next generations won’t have that same ideal to lose
Michał Bortkiewicz
@m_bortkiewicz
Sep 2
We also show that action chunking can work remarkably well in fully online RL, extending its benefits beyond the offline and offline-to-online settings where it has mostly been studied. Check out the work led by @MichalAKorniak and @dybek_kamil !
Kamil Dybek
@dybek_kamil
Sep 2
Action chunking in RL has primarily been studied in offline and offline-to-online settings. Our results show that its benefits extend beyond these regimes: Contrastive RL (CRL), for example, gains substantially from action chunking even in the fully online setting.
Michał Bortkiewicz
@m_bortkiewicz
Sep 2
We provide a new perspective on why action chunking helps: it enables Contrastive RL to learn more informative representations of the states likely to be visited. Check out the paper led by @MichalAKorniak and @dybek_kamil !
Michal Korniak
@MichalAKorniak
Sep 2
Self-supervised RL usually models one action at a time. We extended contrastive RL over action chunks instead, and found large gains: +31.7% offline, +93.1% online. Many explanations exist for why chunking helps. We find a new one: it improves the critic's representations.
Michał Bortkiewicz
@m_bortkiewicz
Jul 11
Reward hacking

Michał Bortkiewicz

Michał Bortkiewicz’s posts