Drafting

[Rob Cashwalker]

Yeah... this should help...

[slowe]

Having a decent pick order for the AI is a good thing, and would help the enjoyability of draft a lot (believe it or not, getting planeswalkers and mass removal eighth pick gets old ). That said, I don't know if gathering data from Forge users is the best way to go about it. I can easily see there being too few different drafters for each format, while a few drafters with a personal bias toward a certain color or archetype could skew the results. Perhaps I'm just too pessimistic...
Regardless, I think using the pick data that have already been aggregated by MagicDraftSim or Bestiaire is a good way to start, even if only until some Forge user data has accumulated. This would provide immediate improvement in AI picks, and the pick orders are much less likely to be skewed because of how many people's picks have been counted. On that note, MagicDraftSim has a lot more picks counted than Bestiaire, particularly for older sets, and seems the better option of the two.

On the other hand, Forge user data leading to a color/archetype becoming "overdrafted" and weaker (because the AI fights you for cards in that color) could actually be a positive. As Accountancy suggested, this could create a metagame where colors wax and wane in viability over time, at least in the most frequently drafted formats. However, it might harm AI drafters a lot more than humans and make draft less interesting. Anyway, enough of my speculation. Thoughts?

[Accountancy]

I was wondering, will cards get more points for being picked first? I also don't think the last pick should be counted at all as there's no choice going into it. Something like first pick gets 17 points, second pick gets 15, third pick 13, fourth pick 11, fifth pick 10, then continuing to go down 1 point a pick till the last pick isn't counted, so you only have to count 14 cards a pack too.

[juzamjedi]

I can easily see there being too few different drafters for each format, while a few drafters with a personal bias toward a certain color or archetype could skew the results. Perhaps I'm just too pessimistic...
...
On the other hand, Forge user data leading to a color/archetype becoming "overdrafted" and weaker (because the AI fights you for cards in that color) could actually be a positive.

I think you changed your mind in the middle of your post Overall yes, The system "should" work so that high human picks train the AI to take those cards early. The feedback you send to Forge DB (Blue control is good, or Red aggro cards, or whatever) should result in you seeing fewer copies of those cards. The "bad" cards will go last and if there's no good use for them in the format they will remain late picks. So Bias in a given format should be self-correcting.

It is my expectation that I will stop seeing Mox last picks. (seriously, this has happened). I draft every Mox I see regardless of archetype. I also expect to see fewer swords because most cube decks run a lot of creatures and sword + creature is a good combination

On a slightly unrelated note: one of the benefits of sending feedback to Forge DB is we will see what cards are consistently drafted late. Cube maintainers (like me, or whoever else) can see which cards are unexciting to people and cut them when we want to add awesome new cards.

[Rob Cashwalker]

I have just added draft pick data collection to the SVN. The default is to collect data, but there's a UI option to disable it, which is saved in the preferences. On startup, it checks http://cardforge.org/draftAI/ping.php if it returns "pong" then the network is available.

[Accountancy]

I was just thinking, how would this affect drafting specific strategies? Like if a lot of players draft Furnace Celebration decks and a lot of players draft infect decks, how will the computer avoid ending up with lots of Flesheater Imps and Artillerizes?

[juzamjedi]

Rob, I was thinking about the draft AI more over the weekend. With pick orders implemented I would suggest changes to the draft AI. Hopefully these suggestions are improvements, if not we can refine it again later.

If first 7 picks: take the "best card". Best card as determined by pick order data collection. (7 is an arbitrary number I just picked, this could be tweaked later). If 2+ cards are tied score, just pick a random out of those 2+.

After first 7 picks: AI evaluates the cards drafted so far to see what are its 2 "best colors" so far. AI will now force these 2 colors, artifacts and lands that go with that color.

1. In a pack, AI will take the highest ranking card that meets the criteria above.
2. If no cards meet criteria, then AI will "hate draft" the highest ranking card.

Deckbuilding: at the end of draft AI should try to build a deck with 23 spells and 17 lands. Start by evaluating the strongest color and try to make a single color deck. Strongest color would be best average among colors with 7+ cards drafted.

If there are not 23 spells then add the next best color with 7+ cards. If you need to add a third or fourth (!) or fifth (!) color they don't have the 7+ requirement.

Finally, add the lands. Start by adding the non-basics that go with color(s) of your deck. Fill out the rest of the mana base by ratio of color multiplied by number of land spots left.

TODO: is there a way for AI to know which lands go with a particular color combo?

[Rob Cashwalker]

Thanks jedi, that's pretty much where I was heading.

The current code, would pick its two target colors based on the strongest cards in its first two picks.
Along your line of thinking, after the 8th pick, your original pack is returned. By then you should have a good sense of where you're going and where other players likely are... hell the AI could even figure out what's missing from the returned packs... maybe....

I like the "hate draft" concept.. but it's really a last resort, as I would rather the AI draft cards from its colors or "good" artifacts....

As long as the AI picked mostly from 2 colors, there's no reason to try forcing a single color deck during build. I don't know anyone who has been able to draft and build a single color deck.

Lands are drafted based on their ability to produce mana of either color the AI has picked. So if the AI is U/B, then it may pick dual lands producing W/U, U/B, B/R, W/B, U/R, U/G or B/G. So if the AI doesn't determine color until pick 7 or 8, then it may miss out on early rare picks for dual lands.. Non-mana producing lands aren't preffered.. yet... unless the rating is high enough to sway the AI's opinion.
I don't have any logical way to check Artifacts for color activation's, or mana production, though the Oracle text addition may help with that... (friarsol - one really important vote for keeping the {} around the mana symbols!)

[friarsol]

I don't have any logical way to check Artifacts for color activation's, or mana production, though the Oracle text addition may help with that... (friarsol - one really important vote for keeping the {} around the mana symbols!)

Ah. That's some fair reasoning. I can comment out the line that's doing the stripping. This may end up creating mana symbols that look different than our current system (specifically with Hybrid Mana Symbols)

[Rob Cashwalker]

Sol, I know things may look odd compared to what we're already doing... but at the same rate, what's really preventing us from "fixing" our current system?

[friarsol]

Sol, I know things may look odd compared to what we're already doing... but at the same rate, what's really preventing us from "fixing" our current system?

Convincing someone to do it :-d

[juzamjedi]

Along your line of thinking, after the 8th pick, your original pack is returned. By then you should have a good sense of where you're going and where other players likely are... hell the AI could even figure out what's missing from the returned packs... maybe....

Now THAT would be awesome The AI can probably be decent without that ability, but figuring out the other colors for the drafters would definitely make the AI better.

I like the "hate draft" concept.. but it's really a last resort, as I would rather the AI draft cards from its colors or "good" artifacts....

Sorry for not being clear, but that was my intention too. Hate drafting is the last resort if there isn't something else that will make your deck.

As long as the AI picked mostly from 2 colors, there's no reason to try forcing a single color deck during build. I don't know anyone who has been able to draft and build a single color deck.

Here is my main point: AI often plays too many colors. It should try to play as few colors as possible to give fewer "free wins" to the human player.

I draft a LOT and I have been able to draft some mono-color decks. They don't happen often but it's certainly possible. So it's actually possible that AI might only want to force 1 color after first 7 picks (let's say Blue). if that happens we could tell the AI to force Blue only. When there isn't a good Blue/artifact/land card it "hate drafts" the best card in the pack. Then at the end of the draft AI looks at its pool and maybe it ends up as Blue/x and the couple of non-Blue cards are amazing bombs.

Lands are drafted based on their ability to produce mana of either color the AI has picked. So if the AI is U/B, then it may pick dual lands producing W/U, U/B, B/R, W/B, U/R, U/G or B/G. So if the AI doesn't determine color until pick 7 or 8, then it may miss out on early rare picks for dual lands..

This seems fine to me. The only strategy that should take duals that high in the draft is 5 color control which we probably don't want AI to attempt. Although I will point out decks will want to take on-color fetchlands too. Do you have a way to check for land type as well as color?

For example, I regularly play Misty Rainforest in my Blue decks even if I'm not running Green. The ability to find your dual lands (Tundra, etc.) is really good.

PS really good discussion, I am very happy to see drafting getting such major overhauls.

[Rob Cashwalker]

Trying to draft fetchlands is pretty narrow - either a rare possibility in a full draft, or a small possibility in 2 blocks, but then there's a pretty good chance in your custom cube....
OK, I guess you can expand that to looking for the Alara Panorama lands....

I think It's just a lot of extra complication. Simply using the pick data should be good enough to bubble the lands to the top of the AI picks...

[juzamjedi]

If coding for land types makes it exceptionally harder then don't sweat it. You're right - there aren't many lands that specify land types without a mana symbol. Mirage has a cycle too (won't ever make it in cube) and Urborg, Tomb of Yawgmoth (which seems unlikely for cube)

[Rob Cashwalker]

If the fetchlands have any AI SVars, then it would be moot anyway. Like your example.. it's a combo-type of land, which enables Swampwalk, and likely a number of other tricks. While the AI certainly could play it functionally, we probably have the flag to not include it in an AI deck because it's too specific.