Come check out our poster tomorrow morning! It will be in Hall C, #1101

Today, Edward Hughes and Michael Dennis from Google DeepMind's Open-Endedness Team will be presenting "Open-Endedness is Essential for Artificial Superhuman Intelligence" as an Oral at 4:30pm in Hall C1-3.

We’re here at 613 in hall C! Stop by to say hi and talk about what open endedness is and how we can achieve it in AI

We have a speaker change: instead of Jack Parker-Holder we'll hear from Michael Dennis - the focus of the talk is the same, though, so join if you're interested in generating any environments you can imagine!

New #RLC2024 paper Three Dogmas of Reinforcement Learning joint w/ Mark Ho and Anna Harutyunyan! arxiv.org/pdf/2407.10583 We reflect on where our scientific paradigm needs adjustment, and suggest three departures from previous conventions. Curious to hear what folks think! 🧵

1 million fps!? looks pretty sick.

On the eve of RL_Conference, some thoughts: 1. If we didn't have this conference, I can't imagine a workshop like Finding The Frame Workshop existing. It's important but just wouldn't fly elsewhere 2. The level of community excitement and involvement has been WILD 1/3

I'm excited to share recent work with Grace Liu and Michael Tang on exploration in RL! A Single Goal is All You Need: Skills and Exploration Emerge from Contrastive RL without Rewards, Demonstrations, or Subgoals Paper, code, and videos: graliuce.github.io/sgcrl/ A thread.

🚀 Excited to share our latest research on combining traditional Procedural Content Generation (PCG) algorithms with Large Language Models (LLMs) to enhance text-controlled game map generation! 🗺

There are many game content generators out there, mostly tailored to specific games. But what if you want to just describe the content you want in words? We present Moonshine: a method for distilling game content generators into text-conditioned generative models!

Cool way of generating interesting environments!

People were dunking on this sort of stuff for being too jank *this year* and the fact we're at this point already is insane x.com/_rockt/status/…

Looks like a cool upcoming venue for people in the open-ended learning space!

My lab had been working on scaling up environment/level generation via reinforcement learning for a while now, and we have some nice results to share with you! PCGRL+, our updated PCGRL framework, is not only an order of magnitude faster but can also scale to larger levels than

People using RL to generate levels for UED should know about PCGRL and PCGRL+ using RL for proc gen more broadly They get more generalizable RL level generators by varying level size, fixing level features, and shrinking its context - would be good to know for PAIRED-like algs!

Turns out prioritising levels that have high learnability leads to pretty good results!

Interesting thoughts on UED regret estimates! There's a tradeoff between promoting generalisation and learnability in UED, regret is a good proxy for both but there are likely even better proxies for learnability I'ld want learnability early in training and generalization late

Should AI be aligned with human preferences, rewards, or utility functions? Excited to finally share a preprint that Micah Carroll Matija Franklin Hal Ashton & I have worked on for almost 2 years, arguing that AI alignment has to move beyond the preference-reward-utility nexus!