Card AI (Improvements) Requests

[martin-cy]

Hi,

Been playing a bit with the new KTK cards..

some behaviour thats a bit questionable:

1. playing Sarkhan, the Dragonspeaker on the second main phase instead of the first one and then doing +1 and attacking for 4 in the air.. thats almost allways the correct play..

2. Butcher of the Horde, when there are tokens on the table it probably make much more sense to play him on first main phase and give haste and lifelink.. sacrificing a 1/1 token to give butcher life link is in most cases worth it..

[lujo]

In the current beta (not the latest snapshot) the AI destroys it's own equipment with the kicked from Kor Sanctifiers.

[slyfox7777777]

The AI makes some obviously terrible decisions in drafting. This mostly pertains to lands. The AI seems to pick even off color lands early, and play them, even if their deck doesn't have the colors. This is really annoying. In KTK and some other sets, the AIs hog all of the fixing lands, regardless of whether they are in their deck's colors, so getting fixing is unrealistically difficult. This is probably the biggest problem with the draft AI. Maybe make the AI think of the lands as colored?

I would really appreciate having the problem fixed, and I'm probably not the only one.
Thanks

[friarsol]

The AI makes some obviously terrible decisions in drafting. This mostly pertains to lands. The AI seems to pick even off color lands early, and play them, even if their deck doesn't have the colors. This is really annoying. In KTK and some other sets, the AIs hog all of the fixing lands, regardless of whether they are in their deck's colors, so getting fixing is unrealistically difficult. This is probably the biggest problem with the draft AI. Maybe make the AI think of the lands as colored?

Watching the Pro Tour this weekend, it sounds like the AI is doing exactly what Ari Lax said he was doing in drafts. Taking the best cards, then taking all the mana fixers so it's more likely that the good cards go around the table more and come back to him.

[lujo]

This is probably the biggest problem with the draft AI.

That's not even close to being the biggest. That happens in any multicolored draft format once stuff gets figured out, manafixing gets snapped up first. Not necessarily because you'll play it, but it leaves your options open among other things. At some point during Ravnica draft bounce lands and signets were really, really high picks.

[ZappaZ]

AI just used a Brimstone Volley on my Knight of the Holy Nimbus without being able to pay the 2 extra mana. (The AI had only 3 lands)

[gecktrix]

Speaking of Draft AI, there were some groundwork laid out and great ideas from this older thread here http://www.slightlymagic.net/forum/viewtopic.php?f=26&t=4817

Was it ever implemented or are there plans to?

[lujo]

The AI plays Smelt-Ward Gatekeepers after combat, which is kinda pointless (especially in case of The Great Gatsby deck).

[lujo]

Speaking of Draft AI, there were some groundwork laid out and great ideas from this older thread here http://www.slightlymagic.net/forum/viewtopic.php?f=26&t=4817

Was it ever implemented or are there plans to?

Speaking of that, it's pretty clear that certain formats and sets simply lack relevant data - the AI is picking up pointless stuff and putting it into it's deck while passing staples in relevant colors. This is especially true for sets which were never drafted or drafted properly on Mtgo.

How does the "upload draft picks" thing work, does it work? Because I don't mind rattling off several hundred drafts for Homelands, Fallen Empires and Masques if that would help straighten out at least some of the lunacies. They're small formats and it's actually pretty easy to sort out the priorities if you actually know the sets well enough.

[gecktrix]

From what I gleamed from the thread (I read the whole thing, really interested in the Draft AI), i do believe that it was implemented where our picks are uploaded to a server, but I do not believe that the AI uses this data, at least not very well.

I would love it if further development was made for limited games for both Draft/Sealed AI, and implemtation of the network code that would allow limited games over the internet.

Not sure if it's possible, but a pretty neat idea would to have the game generate randomized profiles of the AI drafters per draft, where each drafter could have the potential of personality traits which govern which colour preferences that draft perfers, or how likely the AI would splash for a third colour, or how likely the AI would want to stay "open" before sticking to a colour. Seems overly complicated, but it's just a dream.

[lujo]

I'm also very interested in the drafting AI, but I get the feeling that the state it's in currently is very far from where you'd like to take it.

The general attitude towards draft and set implementation isn't helping things - what I get is that most of the older sets were implemented without much enthusiasm and that the developers seem to have thought that noone would be interested in them (and the definition of "older set" could be anything depending on who's talking). They seem to be interested in getting "newer" sets working, but that's kinda missing the point of Forge entirely, as newer sets can easily be drafted IRL, while the first thing any gang of vets would do with forge is jump on a chance to draft stuff that's out of print or which they never got much chance to draft when they started because they were incomless kids (but which we drafted using our collections and home-made boosters, "cube" is MUCH older than people think it is). When properly implemented some of this stuff is actually awesome to draft.

Then there's also the problem that the AI ignores stuff it can't use or doesn't have proper AI for in a draft which leads to staples tableing around against all reason. What I get is that that gets blamed for some AI behavior. Since card AI gets implemented by people personally interested in having a particular card work, many key cards in drafts have likely been ignored because they don't look like something anyone would ever use (so not worth the effort of implementing properly). Which then plays hell on drafts because 95% of mtg cards (if not more) weren't made for tourney or even "constructed" play, and "usable by human only" doesn't mean "implemented" if a set the card is in is meant to be drafted (both in play or in drafting), and that was the criteria by which cards go on the list of "implemented".

So if you have enough archetype defining cards not being drafted, being undervalued or ignored by the AI, you can only call many drafts "working" if you don't in fact have a clue about how they are supposed to be working. I bet a lot of it gets chalked off as "old sets weren't optimized for draft", except most of them were played in and designed to work in an essentially limited environment meaning they're supposed to work BEST in draft. Which is why implementing most cards to be usable by the player in a constructed mode is far less relevant than making the environments (block, block drafts) for these cards work as intended.

But since the whole thing was approached misguidedly draft is a big mess overall, and even trying to discuss it is problematic because fundamental stuff had been ignored and you can get accused of "trolling" or "being too harsh" if you try to bring it up. I had a programmer buddy / cube wiz / veteran mtg player take a look and it and what he had to say about the state of it is basically unprintable, his cube group was mortified/found it hysterical, and unless someone's a competent programmer/coder you can't really breach the subject as something that needs attention.

Even without any fancy work on individual preferences for particular drafters at the table, the current way draft info is pulled is severely lacking and the whole thing is in that worst spot of looking like a real thing but not being close enough to the real thing to not be a dissappointment. And it must've been a lot of work to even get to this point, since people get upset if you criticize it which isn't helping things at all.

[gecktrix]

Honestly, I can see how some might think you are "trolling" or "being too harsh", as it feels like you have unreasonably high expectations for a free program being developed by people in their free time and not getting paid for it. I can see how some of your critisms can be constrewed negatively based on the tone of the comments. Even just based on the last two paragraphs of your reply, I would be hardpressed to work hard on this program if i were a developer.

From someone that has been following this project since '08 (Thank you MTGRares!!!), how far it has come along is insane. We, as users, need to be careful to not have unreasonably huge expectations for this program, as MTG itself is a highly complex game, and not even Wizards themselves can make a fully functional program that doesnt even include AI... as we can see with the debacle of current iteration of Magic Online.

[lujo]

Expecting board sweepers not to table is unreasonably high expectations? You're suggesting personalized preferences for individual drafters, I'm asking for finding ways to fix stuff which anyone who's drafted a set more than once or twice (or just googled top limited commons/bombs for set X) would spot a mile off.

Heck if something is difficult to code in game, I don't mind, the problem is that the holes are so glaring that it's impossible to tell how come whoever left them there, but set out to code mtg draft, could even tell themselves that "it's working". How could you see it like that and leave it like that?

That's basically how a mtg vet would, when playtesting the draft feature, report whether it's working or not - not if the display is working, not if random cards disapear from packs in a dwindling spiral fashin and then you get to play a few matches with suboptimal cards, that's a given, but whether the actual draft is working.

There's no way to say that in a nice tone of voice. The guys who were mortified actually took up my laptop and spun a few drafts. They saw what they were being passed and were shocked - "But you told us they implemented the draft!", and after I tried to give the various excuses they still couldn't figure out who in the world goes out to code up a drafting simulator, gets bombs circling around for 9 turns and concludes that "it's working". No, it bloody well isn't. They didn't care if the AI made suboptimal plays or couldn't dish out fancy moves, but the AI feeding them bombs and making lousy deckbuilding decisions made draft a huge dissapointment, and the general feeling was that that's the one place forge COULD, in theory, bring something to the table.

I mean, you have the AI making tri-color decks to put in stuff that's absolutely worthless under any circumstances, but for whatever reason doesn't at least hate draft, say, Void in Time Spiral draft. Or gives you multiple Dwarven Catapults in FE (set notorios for lack of removal). Or gets data for picks from results of people drafting famous constructed cards which actually suck in draft or are in colors which don't have anything going for them apart from that one thing, feeds you way better picks and doesn't pose a challenge later.

Or the fact that it seems reasonable to randomly set people up with a random inter-block distribution of boosters. Is it unreasonable to expect someone who'd put time and effort into coding a draft simulator to not need the problem with this explained to them? Or their sleeve pulled to maybe get them to consider making it the way it should be?

Why does all this stuff need to be explained at all? I had a group of 9 15-20 year term magic vets looking at me and asking me this the other day, all in disbelief. I asked them would they still use Forge, and the answer was, well, if the draft worked it's be good, but as it is now we got cockoatrice (also named a few other things) if we want to play against each other.

Why does it have to be like that?

[lujo]

On a completely unrelated note to everything:

In the Jamuraa world, there is at least one deck with Frenetic Efreet. The AI just keeps flipping the coin on it all the time. The ability sayes that if it wons, the efreet phases out, and if it loses the efreet dies. It's in the deck/s because that was rather poweful back then as it could dodge removal and board sweepers for free.

If the AI is constantly using the ability, the deck can't do anything at all. Can anyone take a look at it? It would actually be better for that card to never have the ability used than having it used all the time

[Marek14]

On a completely unrelated note to everything:

In the Jamuraa world, there is at least one deck with Frenetic Efreet. The AI just keeps flipping the coin on it all the time. The ability sayes that if it wons, the efreet phases out, and if it loses the efreet dies. It's in the deck/s because that was rather poweful back then as it could dodge removal and board sweepers for free.

If the AI is constantly using the ability, the deck can't do anything at all. Can anyone take a look at it? It would actually be better for that card to never have the ability used than having it used all the time

Maybe you could limit Frenetic Efret to one activation (tell AI "don't activate Frenetic Efreet if its ability is already on stack"). Of course, that ability would need a caveat: "Unless you have Chance Encounter"