Sam Coward

🔧 Looking to easily develop RL novel autocurricula methods? ⚡ Want clean, blazingly fast, and readily-modifiable baselines? Presenting JaxUED: a simple and clean RL autocurricula library written in Jax!

@CsabaSzepesvari @mcbeukman Good catch - you are correct. In our practical implementation, as described in the experiment section, we only actually maintain a single policy, and perform updates wrt the trajectory only after the first k steps (to approximately maintain behaviour on previous adversaries).

@mcbeukman Interesting. Can we simplify the last two lines to: "if tau \not \in \mathcal{T}_\pi( \cup_{j=1}^{i-1} \textrm{Supp}(\Lambda_j)) then" ? (Saving the calculation of k. It seems to me that k=|tau| if and only if tau is in that set which appears in the calculation of k.

Regret is a reasonable objective for unsupervised environment design, right? 🤔 Well, not always, as our new preprint shows. We find that regret does not always correspond to learnability: high regret levels can provide no learning! 1/ 👇

Regret-based environment design can lead to learning stagnation! Exciting work done with @mcbeukman✨

🎮 Introducing the new and improved Policy-Guided Diffusion! Vastly more accurate trajectory generation than autoregressive models, with strong gains in offline RL performance! Plus a ton of new theory and results since our NeurIPS workshop paper... Check it out ⤵️

For more details, visit Arxiv: arxiv.org/abs/2403.13091 GitHub: github.com/DramaCow/jaxued

Special thanks to @MinqiJiang for implementing DCD+minimax, which was a cornerstone and inspiration for JaxUED. Thanks @_chris_lu_ whose hardware-accelerated implementations of RL influenced this work. Thanks to @RobertTLange, whose work inspired JaxUED's environment interface 🚀

🔧 Looking to easily develop RL novel autocurricula methods? ⚡ Want clean, blazingly fast, and readily-modifiable baselines? Presenting JaxUED: a simple and clean RL autocurricula library written in Jax!

I’m excited to announce Craftax, a new benchmark for open-ended RL! ⚔️ Extends the popular Crafter benchmark with Nethack-like dungeons ⚡Implemented entirely in Jax, achieving speedups of over 100x 1/