Drafting

[Rob Cashwalker]

I had actually been thinking about using the human-built deck choices as part of deck building AI... one step at a time though.

[delirimouse]

It might be better to track the AI's success with putting in different cards. There are cards which will be fantastic for humans and bad for the AI because it will not play them well. It would be good if the AI figured out there were some cards it couldn't play well and lowered its tendency to play them accordingly.

[Rob Cashwalker]

You've mentioned trying to make the AI learn from itself... I just don't get how that would help, much less the difficulty in implementing that. In general, we're trying to determine the relative strength of cards compared to each other.

Using pick order to scale the card score was one way I thought might help differentiate each card, not so much that I wanted the AI to pick cards in that order.

[moomarc]

You've mentioned trying to make the AI learn from itself... I just don't get how that would help, much less the difficulty in implementing that. In general, we're trying to determine the relative strength of cards compared to each other.

Perhaps one way that it might be implemented is to add a rating multiplier for AI wins. Not too sure what ratio would actually work, but lets say if the AI wins, the cards in its deck have their rating multiplied by 1.2, or something similar. That way higher ranked cards that actually affect AI wins will be ranked even higher, whereas last picks will only be affected slightly - probably not enough to affect their ranking. This method would slowly build up on player picks to isolate cards that the AI actually uses effectively.

[Sloth]

Not much longer and someone mentions a neural network AI.

[Rob Cashwalker]

yeah, neural network... sure. I'll just whip one up...

It becomes a lot of data to track, but moomarc makes sense.

[delirimouse]

moomarc makes sense.

That is more or less what I had in mind, but I totally understand if it is infeasible. I don't believe it is necessary, but there are a few cards the AI just will not ever play well.

Perhaps a more reasonable approach would be to have drafters manually report such cards and then tell the AI not to play them.

This could be another instance where the light cheating I suggested could come into play. Let the AI replace the cards it can't play with similarly rated cards it can.

I think there is a lot of room to explore letting the AI cheat so it can be a bit more of a challenge.

[Sloth]

Perhaps a more reasonable approach would be to have drafters manually report such cards and then tell the AI not to play them.

We already have a system to tell the AI what cards it can't play well.

[juzamjedi]

In general, I think it's most important to focus on getting the draft pick orders correct and get the AI to draft 2 colors + colorless cards (maybe splashing a 3rd color if they are short on playables). IMO weighting card value by AI performance is a fine idea after the basics get nailed down.

[delirimouse]

Hey Rob, I know you're busy, but if you update the algorithm used for ranking cards can you send a link to the new rankings? Perhaps you can even isolate just the rankings for M12 cards? I'll be very interested to see how closely it aligns with my view of the relative strength of the cards.

[Hellfish]

I'm really looking forward to these improvements, but I will miss the days when I could draft decks like these:

| Open

[metadata]
Name=RB
Deck Type=Draft
[main]
1 Blood Ogre|M12
1 Bloodrage Vampire|M12
1 Chandra, the Firebrand|M12
2 Doom Blade|M12
1 Fireball|M12
3 Flameblast Dragon|M12
3 Goblin Arsonist|M12
3 Goblin Fireslinger|M12
1 Goblin Grenade|M12
1 Gorehorn Minotaurs|M12
1 Inferno Titan|M12
10 Mountain|M12
1 Reverberate|M12
3 Shock|M12
1 Sorin's Vengeance|M12
7 Swamp|M12
[sideboard]
1 Act of Treason|M12
1 Crumbling Colossus|M12
1 Elite Vanguard|M12
3 Fiery Hellhound|M12
20 Forest|M12
1 Goblin Tunneler|M12
1 Gravedigger|M12
1 Greater Basilisk|M12
1 Guardians' Pledge|M12
1 Ice Cage|M12
20 Island|M12
1 Mind Control|M12
10 Mountain|M12
20 Plains|M12
1 Royal Assassin|M12
1 Siege Mastodon|M12
1 Stave Off|M12
13 Swamp|M12
1 Volcanic Dragon|M12
1 Wall of Torches|M12
1 Warpath Ghoul|M12
1 Wring Flesh|M12

[delirimouse]

Yes! Why does that never seem to happen when I draft on MTGO!

[Rob Cashwalker]

FYI, cube draft custom pools decks can use a meta data item, "Custom Pool=true" to prevent the deck from showing up for constructed mode in the new HomeScreen. I have already added this to the juzamjedi draft pool in the SVN.

[juzamjedi]

Awesome, thanks Rob. Now that we have Innistrad (and new cards already in Forge!) I should get around to updating the cube again.

[juzamjedi]

I have noticed that cards repeat now in the draft. This is probably fine for normal draft but not for cube. I have 450 cards in the cube. Is there a setting that I am missing somehow? I have noticed this for a while but just forgot to mention this before.

There are some AI-playable settings that surprise me. For instance I wanted a new UR card. Frenetic Efreet is listed non-playable, but Frenetic Sliver is fine. And we need Sulfuric Vortex in the worst way for Red. At some point I will make a larger list of cards that are not AI-playable that I wish we could add.

Also, cube update attached. I continue to make the cube more aggro-focused including blue.