Maximilian Weichart

📢 Deadline extended!
Submit your work to EWRL — now accepting papers until June 3rd AoE.
This year, we're also offering a fast track for papers accepted at other conferences ⚡

Check the website for all the details: euro-workshop-on-reinforcement-learning.github.io/ewrl18/

Today we concluded our first R1 Reinforcement Learning meetup where I presented and we discussed the paper on AssistanceZero (by @cassidylaidlaw.bsky.social et al.)

If you're interesting in joining & talking about RL check out the meetup 💡 max-we.github.io/R1/

Super interesting, I didn't know Minecraft could be used in such a way as a GCRL training environment! I'm wondering if one could also extend the agent to use the in-game chat and how that would change the architecture.

Although 5000+ participants (and counting!) after just 30 minutes would be wonderful, I switched out the form in favor of fillout.com, as it seems like Google-Forms has a serious bot problem.

Interested in RL? I'm planning to assemble a new online meetup, focused on reinforcement learning paper discussions. You can sign up, and as soon as enough people are interested, you'll get an invitation.

More information and registration: max-we.github.io/R1/

GitHub - Max-We/alphazero-tetris: An implementation of AlphaZero and MCTS with neural networks for Tetris An implementation of AlphaZero and MCTS with neural networks for Tetris - Max-We/alphazero-tetris

Open-sourced my implementation of AlphaZero and various other MCTS policies to play Tetris. In contrast to other Tetris-agents, this implementation does *not* rely on observation- or action-space simplification. It trains an agent with the same information a human has.

github.com/Max-We/alpha...

Wenn im Endeffekt durch ein solches Format die Kompetenzen eines LLMs statt der Studierenden überprüft wird sollte man sich tatsächlich Gedanken machen

It's always been part of their marketing and certainly doesnt help anyone in the long run...

Here's a more honest picture with 25 evaluation games (training lasted 3 days, but can be scaled up a lot more!)

🥳 50k score achieved, TetrisZero is working! Here's the viewer-site with a replay (actually, the replay became so long that the site is lagging a bit, lol). Full details on the algorithm will follow, once I evaluate it against AlphaZero...

max-we.github.io/tetris-zero/

It's getting there! Target is a score of 50k, currently about 10k.

W&B sweeps is a really nice way of hyperparameter-searching. Didn't see a lot of people talk about it, but it makes the process really nicely streamlined + visualized. Essentially, you just need a config-file with the parameters to try, and it's ready to go

docs.wandb.ai/guides/sweeps/

I found this via Scholar Inbox today. These are detailed, clear and understandable explanations + exercises to learn with. Thank you, great work!

Inspired by the interface that DeepMind used for AlphaGo (sadly closed source, at least I couldn't find anything on the web about it)

Working on debugging RL algorithms such as AlphaZero is hard, especially when the codebase uses just-in-time-compiled JAX. So I created a replay-viewer which visualizes an episode with all the policy statistics for a personal project. Will be open-sourced once I finish my new algorithm!

finally!

Is it normal to thank ChatGPT in the Acknowledgements of your paper nowadays? lol

arxiv.org/pdf/2301.01379

ps.is.mpg.de/pages/princi...

This is a really nice alternative (or even better service) for following research papers than _akhaliq on X, nice!

Excited to share that today our paper recommender platform www.scholar-inbox.com has reached 20k users! We hope to reach 100k by the end of the year.. Lots of new features are being worked on currently and rolled out soon.