• Follow our Instagram!
  • Curious about what we're currently doing about bots? Check out the latest updates from staff in this thread.

Battle Bot Pokerus: Selfplay + Teambuilder Bot reached #1 on Gen9 OU ladder

Hi everyone!

I'm a researcher at MIT and an avid Pokemon fan. I've been working on a research project lately involving reinforcement learning and teambuilding for Gen 9 OU.

I was recently testing my Pokemon Showdown harness under the username "fluttermanemain" and was pleasantly surprised by the results so far, so I thought I'd share a bit about the project. The bot currently uses a simple policy with no search, taking around 0.2 seconds per decision. The bot has generated over 50 million teams through self-play so far, and training is still ongoing.

My eventual goal is to test the bot against some of the strongest human players in the format, once training and the search algorithm are ready. With more bots appearing on the ladder, it's becoming harder to gauge performance from ladder results alone.

I know there's been a lot of discussion here recently about bots and their impact on competitive Pokemon. I don't currently have any plans to release the bot or its code, and I'd like to be mindful of the community's concerns. Mostly, I wanted to share what I've been working on, gauge interest in the project, and hear people's thoughts.

Sample team: https://pokepast.es/9f2601a00e667d95

Also, shoutout to Xatu Zero for being the first to reach #1!

1793853073130.png
 
I was recently testing my Pokemon Showdown harness under the username "fluttermanemain" and was pleasantly surprised by the results so far, so I thought I'd share a bit about the project. The bot currently uses a simple policy with no search, taking around 0.2 seconds per decision. The bot has generated over 50 million teams through self-play so far, and training is still ongoing.
Teambuilding is definitely a new frontier for Pokemon AI. How does the model give out rewards of a team that it has built?
 
Back
Top