In 2023, my friend and I resolved to do a Pokemon challenge in the National Dex Monotype format. We each drafted 9 of the 18 types, and then we drafted 54 Pokemon (enough to fill exactly 9 teams, if there were no overlap).
We used "snake order" for all drafts.
I drafted the following types:
Electric
Bug
Steel
Fire
Flying
Ground
Ghost
Fighting
Ice
He drafted the following types:
Water
Dragon
Psychic
Poison
Normal
Fairy
Grass
Dark
Rock
Here are the Pokemon we drafted. I don't know if there have been any relevant changes to the ban list since, but if so, I will update later. Dual-type Pokemon can be used on both of their respective teams.
We did blind picks for the first matchup, and it was decided that it would be my Fire team versus his Rock team. For your benefit, he has spent a significant amount of time piloting his Rock team, whereas I spread my testing time much more evenly across all 9 teams. He peaked at 1491 Elo, which was in the top 500.
We never got around to doing the challenge and eventually lost interest. However, recently I've been experimenting with programming Pokemon AI, and I've asked my friend to resume the challenge with me so that I can use it as the basis for my project. I will be training an agent in Poke-env for each matchup.
For the first round, the final version of each team will be submitted blind, and then after some amount of time to train, the set will be played.
All rounds will be best of 5. The winner of the round is locked into their type choice, and the loser of the round gets to counterpick their next type. Both players submit their next team (again blind), and the process repeats.
Resolves to YES if my AI wins the challenge (eliminates all 9 of my friend's teams).
Resolves to NO if my AI loses the challenge (has all 9 of its teams eliminated) or if it fails to win the challenge before January 1st, 2026 (deadline may be extended if delays are caused by my friend rather than me).
Resolves N/A if the challenge can no longer proceed for reasons outside of my control (e.g. my friend cannot or will not participate).
I will be participating in the market, but I will only buy YES shares, and I will never sell them. Replays will be posted after each round.
Update 2025-16-01 (PST) (AI summary of creator comment): Details Clarified:
Counterpicks are chosen by the human players, not the AI.
1,000
3.00I've made a bit of progress but I've hit two big walls. Using existing examples, I've got a program that trains a DQN versus a SimpleHeuristics player. Preliminary testing suggests that the Fire team is favoured versus the Rock team (after training each DQN for 100000 steps, the Fire DQN won 39 out of 50 games versus the SimpleHeuristics player, whereas the Rock DQN won only 3).
Here are the two problems:
There are plenty of examples showing how to get a SimpleRL player training versus other defined player classes (random, max damage, simple heuristics), but I'm having difficulty getting it to train versus another DQN player. I'm going to have to experiment with this to see if I can get it working.
In the existing examples I've found, the embedding is very barebones, and I will need to expand it significantly, and I'm not sure how hard it's going to be to teach myself how to do it properly.
Stay tuned for more progress.