Read on Twitter
Finally found the time to finish tuning my MTG Draft AI. Will try and find time to write the code that lets it read draft logs in realtime so I can stream, but the initial results look promising. Looking into the network, here is how it understands top cards in archetypes. 1/X
But note that the model understands nuance. it's clear that the model learned that UR is draw-two and WG is adventures. But again, it still understands that Wicked Wolf > Wandermare in WG, but it doesn't have to rank wolf as highly in WG because it is ranked highly elsewhere. 2/X
The model attempts to create a wholistic understanding of each archetype. And then, a pick is determined by those understandings alongside the bias that the current pool pushes towards each archetype. It learned to over-rank gold cards within archetypes because they are worth 3/x
nothing outside of that specific archetype. The same goes for synergy cards like Mad Ratter. In archetypes that lack synergistic depth, it learns basic synergies (e.g. you can see knights somewhat matters for BW), but will just rank the best cards there the highest instead. 4/X
To test how this holds regarding archetypal bias, I decided to look at the top3 uncommons for P1P1 according to the model. They are as follows:
1 - Syr Konrad, the Grim
2 - Clockwork Servant
3 - Epic Downfall
These are actually my personal top3 uncommons in the same order!
5/X
1 - Syr Konrad, the Grim
2 - Clockwork Servant
3 - Epic Downfall
These are actually my personal top3 uncommons in the same order!
5/X
Lastly, if you consider the model's 1st and 2nd choice vs a humans 1st choice, accuracy is ~91% (75% on first choice only). I will note this is *not* a proper validation metric, so take it with a grain of salt, but it shows promise and I'm excited to stream it!
6/6
6/6
Oh, also because I can throw a little shade, check out the model's "understanding" of UB. Generally good cards/removal . . . BUT IT TAKES MERFOLK SECRETKEEPER HIGHLY 
