I took interest in VGC bots when I saw mikumiku hit rank 1 on the ladder.
I am a pretty big noob in VGC, all things considered, (peak Masterball 3), but I do have a Computer Science degree being wasted away as I type into claude all day at work. So I decided to see what I could make myself, and I reached Top 20 on the ladder (peak #16)
What I Did:
RotomDotExe learned mostly by playing itself, about 139 million games over 4 days. That plateaued and I wasn't sure how to get further only using self play, so to break through I downloaded 6,000 human games from Showdown replays and trained a second model to predict what real players do.
The bot now uses that prediction to look a couple of turns ahead and it plays out its best options against the moves a human is most likely to make, then picks the one that holds up best. Its own ladder games feed back in, too, so the predictions keep getting better. That alone took it from around 1600 to 1800+ on the ladder.
What I want to Do:
I think that VGC bots have a lot of potential, and I am working on is a way to
1: Test a team - the bot pilots your team against every team in a database of recent tournament teams and shows what it’s good and bad against, plus which of your Pokémon are pulling their weight. This can fail if the bot has not used a pokemon on your team very much, and figuring out how to train that in is a step im working on now.
2: Help Create a team - Submit it 1-5 pokemon and allow it to fill in the remaining pieces to see how it can perform the best. This is definitely the most complicated part.
3: Create a game plan sheet - Listing out the common teams, and figuring out 1 - 2 gameplans that the bot would do in order for a new player to understand the team more
Why am I posting this?
I also want to make this post to show it really is not that hard to create a bot that is good at beating humans. I have almost 0 machine learning experience before this experiment, and am normally a front end dev. It's awesome I can produce this, but also shows that many people can, especially the bad actors. I hope showdown and online tourneys can implement a way to crack down on people that would use this to cheat in tournaments. I also hope showdown realizes how easy it is to make these bots and puts in some prevention to playing these on ladder
(I played about 400 ladder games to get a judgement of where the bot is at, but have since turned it off)
I am a pretty big noob in VGC, all things considered, (peak Masterball 3), but I do have a Computer Science degree being wasted away as I type into claude all day at work. So I decided to see what I could make myself, and I reached Top 20 on the ladder (peak #16)
What I Did:
RotomDotExe learned mostly by playing itself, about 139 million games over 4 days. That plateaued and I wasn't sure how to get further only using self play, so to break through I downloaded 6,000 human games from Showdown replays and trained a second model to predict what real players do.
The bot now uses that prediction to look a couple of turns ahead and it plays out its best options against the moves a human is most likely to make, then picks the one that holds up best. Its own ladder games feed back in, too, so the predictions keep getting better. That alone took it from around 1600 to 1800+ on the ladder.
What I want to Do:
I think that VGC bots have a lot of potential, and I am working on is a way to
1: Test a team - the bot pilots your team against every team in a database of recent tournament teams and shows what it’s good and bad against, plus which of your Pokémon are pulling their weight. This can fail if the bot has not used a pokemon on your team very much, and figuring out how to train that in is a step im working on now.
2: Help Create a team - Submit it 1-5 pokemon and allow it to fill in the remaining pieces to see how it can perform the best. This is definitely the most complicated part.
3: Create a game plan sheet - Listing out the common teams, and figuring out 1 - 2 gameplans that the bot would do in order for a new player to understand the team more
Why am I posting this?
I also want to make this post to show it really is not that hard to create a bot that is good at beating humans. I have almost 0 machine learning experience before this experiment, and am normally a front end dev. It's awesome I can produce this, but also shows that many people can, especially the bad actors. I hope showdown and online tourneys can implement a way to crack down on people that would use this to cheat in tournaments. I also hope showdown realizes how easy it is to make these bots and puts in some prevention to playing these on ladder
(I played about 400 ladder games to get a judgement of where the bot is at, but have since turned it off)
Last edited:
