Programming Machine Learning Teambuilder

[MMC]

Hey y'all.

I'm attempting to train a ML network to learn how to create "good" pokemon teams. The usage stats from showdown are helpful, but what I really need is a large archive/database of teams with which to train the network (which as far as I can tell is not included anywhere in https://www.smogon.com/stats ). I was wondering if anyone was aware of any such databases (or alternatively they'd be interested in creating one with me in the event that no such database exists).

 

[zben]

Hello, I want to know what will be your train dataset.
1. good team vs Bad team, binary Classification
2. Ranking Classification, team A better Than team B
3.an other approach, like forcing 1 or 2 slot and create a model that find the best team with this a priori.

1 is not that simple because you need also bad teams dataset, and I think the model will just fit the usage statistics.
The second one will be more interesting but you need a huge dataset of battle result.

 

[Maxouille]

Here you are https://www.smogon.com/forums/threads/smogon-rmt-team-dump.3622884/

 

[MMC]

Here you are https://www.smogon.com/forums/threads/smogon-rmt-team-dump.3622884/

Bless tysm

 

[MMC]

Hello, I want to know what will be your train dataset.
1. good team vs Bad team, binary Classification
2. Ranking Classification, team A better Than team B
3.an other approach, like forcing 1 or 2 slot and create a model that find the best team with this a priori.

1 is not that simple because you need also bad teams dataset, and I think the model will just fit the usage statistics.
The second one will be more interesting but you need a huge dataset of battle result.

Yeah I should've been more specific. Basically the answer is whichever I can get. I should be able to make it work with either approach 1 or 2. While approach 2 would be easier to work with, it also seems like it'd be more difficult to get. It'd require some way of quantifying a good or bad team, which is the aim of this project in the first place so it seems a little dubious. One potential idea would be to weight the teams by the ELO of the players that submit them and combine that with how well they match usage stats but I'm not entirely sold on that method.
As for approach 1, my idea would be to just make the assumption that a user created team is "good" or at least good enough. Given a substantial enough dataset bad teams will be diluted out assuming that in general, people know how to make an decent team. Bad teams could then be generated by modifying submitted teams and randomly changing 2-3 pokemon to through off the balance. It would require some work but it would be much more straightforward than what I'd have to do for approach 2.

 

[zben]

What will be your model architecture and your features , tabular data with pok1 pok2.. Pok6, item 1 item 2... Item6 etc.... ?
I think the best way to use tabular data is to consider the set of a pokémon and not the pokémon itself.In smogon you can find the best sets of each pokémon. The same pokémon with 2 different set should be considered as 2 different pokémon.

You can also try to use an embedding approach like Mango Smoothie in this Poke2vec

 

[MMC]

What will be your model architecture and your features , tabular data with pok1 pok2.. Pok6, item 1 item 2... Item6 etc.... ?
I think the best way to use tabular data is to consider the set of a pokémon and not the pokémon itself.In smogon you can find the best sets of each pokémon. The same pokémon with 2 different set should be considered as 2 different pokémon.

You can also try to use an embedding approach like Mango Smoothie in this Poke2vec

Sorry for the delayed response. I had to step away and deal with finishing up my classes and am now getting back to working on this project. I was planning on doing tabular data but I'm not sure that using it in the manner you suggest would be optimal. While it would pose benefits to storing different sets as different pokemon, the algorithm that would result would then only be able to reuse pokemon sets that have already been created. I worry that the program would only be able to match new pairs of pokemon but would struggle with innovating new sets as it wouldn't know that 1 pokemon could have multiple sets.