Preamble
Niche formats on Pokémon Showdown have a fundamental problem: you can't find games.
Scheduled roomtours, @matchmaking roles on discord, and team tour teammates only go so far. Gen X XU could be someone's favorite metagame, but they'd never know it because it is too awkward, cumbersome, tedious, or difficult to actually a get a game in the format.
The fundamental jump point in making a metagame accessible is the ability to get a game on-demand.
If you want a game, can you get a game immediately?
Competent but not necessarily superhuman bot play is a valuable asset for metagames that do not meet the critical mass required for on-demand play. Obviously, having access to steady human play is ideal. But in the absence of that, access to bot play is useful in creating both interest in a metagame and a sounding board for people who want to test teams.
Paradoxically, by creating bots for a niche human metagame, you actually create an increase in human players of that metagame.
Introduction
Tactigon is a self-play bot that plays exclusively BW LC, a niche format, at a competent level. It currently accepts challenges as Tactigon on Pokemon Showdown and you can view its publicly played games at https://tactigon-bot.github.io/ Privately played games (check the Don't Allow Spectators button) are not saved publicly. Training is ongoing and it is NOT playing at its strongest checkpoint, it is intentionally just strong enough to feel capable without feeling frustrating.
I would rate Tactigon as... good. It plays solidly and tenaciously, but not unbeatably so. In many ways, this is good for testing; you have to be disciplined throughout the game as it will squeeze you for even a chance of winning and if your team is bad you will get teamchecked by the bot using a standard team and playing it consistently well. The goal of this project is not necessarily to advance the frontier in terms of raw strength, but to advance it in terms of cost, scalability, and humanness. When playing this bot, do you feel that this was time well spent and that you learnt something about your team or that you improved as a player? This is a more important question than whether or not it felt unbeatable (although the two are often correlated regardless). Then, can we replicate, generalize, or otherwise extend the findings here for any format and train it as cheaply as possible?
I selected BWLC because
Technical Details
Learning Algorithm and Play
The highest priority teams are teams that from sample teams, Smogon dumps, and other places on the Smogon Forums. After that, I scraped every tournament replay in the recorded history of BWLC and filled in unrevealed movesets and EV spreads programmatically from Smogon analyses and usage stats. Finally, I automatically built teams from the Pokemon Showdown usage stats. The latter two are obviously flawed for many, many reasons. But, that is an acceptable enough tradeoff in order to test vs a meaningful quantity and diversity of teams. I've also weighted the data to focus more on curated teams and teams that perform well within the internal 'tournaments' that get run as part of training.
This is very much a Baby's First Bot and through the process I have made (and am probably still making!) mistakes and inefficiencies. Nothing here is particularly novel, mainly just applying existing techniques and methods to a new format. Treat this as just another data point in the sample set of outcomes.
Future Work
Model strength is inextricably linked with a model's usefulness, so that remains one of the focuses of future work. Any auxillary tool that you build around a model like a match analyzer, matchup analyzer, team rater, etc all fundamentally depend on the model being strong enough to be treated as a divine source of truth. For example, I gathered the top teams that this bot rates as its best from its self-play. Notably, it doesn't have a single BW Hippopotas Sand Team in it's Top 10, despite this archetype being thought of as a good archetype. Is that because we are actually overvaluing Sand, or that the bot is just too weak to understand the long term play that Hippopotas-Lileep teams require?
Also, the second goal is to answer the first Open Question listed below.
Open Questions
Do you want a bot for your metagame? If you are reading this, message me on discord or Smogon if you are interested in setting up a bot for your particular format. Unfortunately, training and server costs aren't free, but it's useful to gauge interest and I'll see what I can do :)
References, Citations and Other Reading:
https://pkmn.ai/
https://github.com/pkmn/teams
https://github.com/HibikiDev/openVGCengine
https://jaxcalibur.github.io/
https://www.smogon.com/forums/threads/nessie123-an-ots-vgc-bot-that-topped-the-reg-m-c-bo3-ladder.3789213/
https://www.smogon.com/forums/threa...the-vgc-reg-m-c-ladder.3789199/#post-11178569
https://openai.com/index/competitive-self-play/
https://pmariglia.github.io/posts/foul-play/
https://arxiv.org/html/2601.03306v1
Niche formats on Pokémon Showdown have a fundamental problem: you can't find games.
Scheduled roomtours, @matchmaking roles on discord, and team tour teammates only go so far. Gen X XU could be someone's favorite metagame, but they'd never know it because it is too awkward, cumbersome, tedious, or difficult to actually a get a game in the format.
The fundamental jump point in making a metagame accessible is the ability to get a game on-demand.
If you want a game, can you get a game immediately?
Competent but not necessarily superhuman bot play is a valuable asset for metagames that do not meet the critical mass required for on-demand play. Obviously, having access to steady human play is ideal. But in the absence of that, access to bot play is useful in creating both interest in a metagame and a sounding board for people who want to test teams.
Paradoxically, by creating bots for a niche human metagame, you actually create an increase in human players of that metagame.
Introduction
Tactigon is a self-play bot that plays exclusively BW LC, a niche format, at a competent level. It currently accepts challenges as Tactigon on Pokemon Showdown and you can view its publicly played games at https://tactigon-bot.github.io/ Privately played games (check the Don't Allow Spectators button) are not saved publicly. Training is ongoing and it is NOT playing at its strongest checkpoint, it is intentionally just strong enough to feel capable without feeling frustrating.
I would rate Tactigon as... good. It plays solidly and tenaciously, but not unbeatably so. In many ways, this is good for testing; you have to be disciplined throughout the game as it will squeeze you for even a chance of winning and if your team is bad you will get teamchecked by the bot using a standard team and playing it consistently well. The goal of this project is not necessarily to advance the frontier in terms of raw strength, but to advance it in terms of cost, scalability, and humanness. When playing this bot, do you feel that this was time well spent and that you learnt something about your team or that you improved as a player? This is a more important question than whether or not it felt unbeatable (although the two are often correlated regardless). Then, can we replicate, generalize, or otherwise extend the findings here for any format and train it as cheaply as possible?
I selected BWLC because
- My own domain knowledge of the format to help diagnose issues
- LC games tend to be shorter so there's less of an emphasis on very-long-term thinking and stall
- Absence of complicating mechanics like Z-Moves or Terastalization.
- vibes, realistically a number of metagames fit this criteria, I just happened to pick this one
- Difficulty in evaluating strength externally - there is no public ladder and there are unlikely to be other bots to benchmark against
- Difficulty in getting teams - There are far fewer publicly available teams, and this is exacerbated if there's ever a ban in the format. Teams can be created by scraping public tournament replays and inferring missing moves and EV spreads, or even by generating teams from usage statistics, but these teams tend to be lower in quality compared to complete teams found on e.g a Smogon Team Dump.
| Opponent Name | Win-Rate of Tactigon vs the opponent |
| Greedy Bot (uses the highest damage move every turn) | 96.1% ±2.4 |
| Foul Play Bot | 0.738 ±0.043 |
| Freezai | 0.55 |
| Miscellaneous LC Tournament Players | 0,67 |
Technical Details
Learning Algorithm and Play
- Self-play PPO, played itself in BW LC, ~50,000,000 times
- Pure Policy, no use of search. A search add-on the the net is WIP but search while training was too expensive for my immediate goals
- Always picks what it evaluates as the strongest move. This is the most desirable (is it?) for someone who's using it to train.
- Experimentally, if you are playing exploitatively, i.e. you are taking advantage of the fact that it always picks the 'strongest' move, the bot flags you as an Exploitative Player and switches to a weighted average of its top moves.
- 7.5M Parameters, pretty standard. The model is small enough that:
- In theory, you *could* get this even smaller with not much loss in performance, but I see no real benefit in distill or quantizing the net further. I haven't investigated further.
- CPU inference works fine.
- I took HibikiVGC's open source rust-based VGC simulator because it had Gens 5 to 9 and I one-shot a singles version. No real dev effort was spent on this beyond fuzzing the doubles to singles changes.
- About 250x faster than Showdown
- ~$45 of compute using a 4090 that I rented using vast.ai to train the model
- Serving the model with CPU at negligible cost
| Team Source | # of teams (unweighted) |
| Scraped from replays | 10,947 |
| Generated from usage stats | 300 |
| Curated teams from team dumps/sample teams/etc | 16 |
The highest priority teams are teams that from sample teams, Smogon dumps, and other places on the Smogon Forums. After that, I scraped every tournament replay in the recorded history of BWLC and filled in unrevealed movesets and EV spreads programmatically from Smogon analyses and usage stats. Finally, I automatically built teams from the Pokemon Showdown usage stats. The latter two are obviously flawed for many, many reasons. But, that is an acceptable enough tradeoff in order to test vs a meaningful quantity and diversity of teams. I've also weighted the data to focus more on curated teams and teams that perform well within the internal 'tournaments' that get run as part of training.
This is very much a Baby's First Bot and through the process I have made (and am probably still making!) mistakes and inefficiencies. Nothing here is particularly novel, mainly just applying existing techniques and methods to a new format. Treat this as just another data point in the sample set of outcomes.
Future Work
Model strength is inextricably linked with a model's usefulness, so that remains one of the focuses of future work. Any auxillary tool that you build around a model like a match analyzer, matchup analyzer, team rater, etc all fundamentally depend on the model being strong enough to be treated as a divine source of truth. For example, I gathered the top teams that this bot rates as its best from its self-play. Notably, it doesn't have a single BW Hippopotas Sand Team in it's Top 10, despite this archetype being thought of as a good archetype. Is that because we are actually overvaluing Sand, or that the bot is just too weak to understand the long term play that Hippopotas-Lileep teams require?
Also, the second goal is to answer the first Open Question listed below.
Open Questions
- What is the best approach to extend this to the training of Gen X XU? Training an overall model for a generation and then finetuning on an individual format? Will that reach the performance necessary? Or is it better to train a model from scratch for each format and simply do that as cheaply as possible? Or is it to become search-focused instead? BW LC is just the initial proof of concept for a niche format bot, how do we scale this to as many metagames as possible?
- Open Source. Inherently, whoever owns a bot has access to all the private replays that the bot plays; I have access to all the private BW LC games that Tactigon plays. The only real solution to this is having the ability to deploy your own bot locally that way no one else can look into it. But this obviously runs into the whole cheating problem.
- Playing more 'human'. Right now, the public version of Tactigon is running a weaker checkpoint without Search so it doesn't feel too difficult to defeat. It is easy to implement difficulty by uploading several different checkpoints of varying skill, but these weaker checkpoints are not fun to play against. They play normally but make catastrophic blunders and those wins aren't satisfying. "Humanness" is known to be a tough problem to solve but there is research in the field, like with https://www.maiachess.com/ which requried a lot of human data as training.
Do you want a bot for your metagame? If you are reading this, message me on discord or Smogon if you are interested in setting up a bot for your particular format. Unfortunately, training and server costs aren't free, but it's useful to gauge interest and I'll see what I can do :)
References, Citations and Other Reading:
https://github.com/pkmn/teams
https://github.com/HibikiDev/openVGCengine
https://jaxcalibur.github.io/
https://www.smogon.com/forums/threads/nessie123-an-ots-vgc-bot-that-topped-the-reg-m-c-bo3-ladder.3789213/
https://www.smogon.com/forums/threa...the-vgc-reg-m-c-ladder.3789199/#post-11178569
https://openai.com/index/competitive-self-play/
https://pmariglia.github.io/posts/foul-play/
https://arxiv.org/html/2601.03306v1



