Over the last couple years I've been training a reinforcement learning agent to play pokemon red. I put together a video documenting the AI's learning, and my process creating it. Enjoy!
🔗 youtu.be/DcYLT37ImBY
Join us on August 13 for Localhost!
🦠 @computerender will present Mote, an interactive ecosystem simulation with hundreds of thousands of organisms. His custom GPU physics engine models many simple behaviors at a massive scale, producing fascinating emergent phenomena.
RSVP⬇️
🦠Watch @computerender's first public talk on Mote: an interactive ecosystem simulation!
Mote uses a custom GPU-based physics engine to model hundreds of thousands of organisms, leading to fascinating emergent phenomena: a sandbox that is part game, part research.
Link below⬇️
Amazing to see how quickly a community is forming around RL for Pokemon. @dvruette and others have been working on support for gen3 games and are already making a ton of progress! More soon
Baby steps... AI has previously learned to play Pokemon Red, and now it's learning to play Pokemon Emerald as well. What will be the biggest obstacles on its journey to become a Pokemon master?
Excited to finally share our progress in developing a reinforcement learning system to beat Pokémon Red. Our system successfully completes the game using a policy under 10M parameters, PPO, and a few novel techniques.
Blog posted below
Peter Whidden @computerender will be presenting his work on playing Pokemon with reinforcement learning at LIFE monthly Monday 11/13 at 9 AM PT / Noon ET. DM me your email for an invitation!
The work Mykhailo has been doing on this project is incredibly impressive. He's created the first tensor/ml library that supports pytorch nn style code, realtime rendering, and non-nvidia gpus. He's already made a ton of amazing demos. Check out the blog post & star the repo!
I wrote a blog post about my own Python tensor compiler library called TensorFrost that I've been working on for the last 14 months!
michaelmoroz.github.io/WritingAnOptim…
As a sneak peak here is the last pet project that I've made using it - an implementation of Neural Cellular Automata:
Can dynamics be distilled from models trained on only static 2D images?
An experiment distilling Stable Diffusion into @zzznah's Neural Cellular Automata model using @poolio's method from DreamFusion.
Prompt: "raindrops on glass"
More info and code: github.com/computerender/…
Given this, perhaps it is possible that dynamics could be distilled into a model that enforces a consistent structure over time and space. These experiments are a first step in this direction!
