VGC team generation
In progress · Paper in preparation
Quality-diversity search for Pokémon doubles teams that are strong and different from each other.
Problem
In competitive Pokémon doubles (VGC), players bring a team of six under a rule set that changes every few months, and the number of legal teams is enormous. Most players start from teams that already win. We want a search that returns many strong teams that play in different ways, so players can find good ideas outside the current meta.
Done so far
- A battle simulator for scoring generated teams, so each candidate can be played against current meta teams many times.
- A pilot evaluation on VGC-Bench that scores a team by its win rate against Regulation M-B meta teams, weighted by how often each meta team is used.
Approach
Quality-diversity search
MAP-Elites keeps an archive of teams sorted into bins by features that describe how a team plays, and keeps only the best team in each bin. The result is a spread of strong teams with different styles, not a single winner.
Scored by playing, not by a formula
Each candidate is scored by simulated battles against the teams people actually use, weighted by usage, instead of by a hand-written rating of the team.
How it fits together
- Legal Pokémon pool for the format
- Team generator + MAP-Elites archive
- Battle simulator
- Usage-weighted win rate
- Archive of strong, varied teams
Status
- No results yet. The search has not been run at scale, so there is nothing to report on team quality.
- Win rate depends on the battle agent that plays the teams, so it measures a team as that agent plays it.
What's next
- Finish a faster battle engine in Rust, a port of Pokémon Showdown's rules, for the full Regulation M-B pool, and connect it to the evaluation.
- Run MAP-Elites over full teams and report quality and diversity.
- Finish the paper. The arXiv link goes here once it is posted.
Credits
Built with Bryce Martin as a Computational Creativity class project.