Card AI (Improvements) Requests

[Fizanko]

Another idea would be for cards like Avacyn, Guardian Angel, let them still be drafted and used in decks, but restrict the AI from using the activated abilities that give it problems. That way it will still have a 5/4 flier for 5, which is a bomb by itself.

That sounds like a very good idea, some AI restricted cards could still be usable even without their AI-unfriendly abilities in term of P/T vs mana cost, if the restriction was only on the abilities it could still allow inclusion of such card in an AI deck to fill a specific spot.

[lujo]

^ Also a VERY good idea. There's a ton of old cards and sets which would highly benefit from this approach too, as it's usually the useless tacked on and marginal abilities on a card which has perfectly sound stats and pick order that are causing problems. It would do drafts in general a world of good.

The way it is now, the AI is effectively just hate drafting stuff, which probably also gets it interested in a color, and the deck ends up being some 4 colored nonsense. A prime example of this is Necrite from Fallen Empires which the AI can even use, but doesn't put in the deck (maybe because it's not very good at choosing targets), but just having it in a deck would be very good for it as it's stats are fine and it's creature type is highly relevant in the set. This alone would solve a bunch of problems with that particular draft.

Another example from the same set is Vodalian War Machine - the AI doens't really need to ever use its abilities, just having a 0/4 wall for 1UU is big in the format until it can cast it's big monsters. If it only ever pumps it after blockers are declared - even better, it can't gimp itself that way! I could go on and on XD

[CyanKnight]

Hello, first post on these boards.
For now I'm using Forge to play Momir, so I think I will encounter a lot of unusual cards.

I did not see a list of know issues, so I assume it's ok to post whatever I find?

Right now the AI has a Spike Soldier out and it keeps moving a maximum amount of tokens from Spike Soldier to (the same) Spike Soldier. It's not very effective. (Same goes for other Spikes, I just encountered Spike Hatcher and it has the same problem.)

AI controlling an Empyrial Archangel will let an attacker through, even if the attacker has enough power to kill the Angel and low enough tougness to trade with the Angel were he to block it.

It seems that the AI values creatures that are borrowed with Act of Treason -like effects as if they would stay on your side.

AI opponent just removed his entire deck with Trench Gorger. I get that it is often a good idea to remove all lands when you can cast Trench Gorger, but maybe it could have a clause to leave at least one card in the library, or more if there are mill/draw effects on the board.

In the Devour/tokens deck the AI makes some notable mistakes regarding Goblin Assault.
First off, he plays the Devour creatures after combat, even if the goblins (that have to attack) run themselves into blockers in the combat phase.
Also, he is stacking the upkeep triggers in the wrong order when he has a Dragon Broodmother in play. He puts the Goblin Assault trigger on the stack first, so the Devouring dragon comes into play before the Goblin token that it would like to devour.

In the anti-black deck, the AI uses Northern Paladin before casting Compost

[gecktrix]

Another question regarding the Draft AI... Pack 1 Pick 2, M14, got passed a foil Archangel of Thune which is #3 on the rankings list. It was uncommon missing from the pack.

Does this mean that the current draft AI determines what colour it wants to be in first before the draft even starts? Or does it mean that the foils aren't ranked by the AI?

[Sloth]

Another question regarding the Draft AI... Pack 1 Pick 2, M14, got passed a foil Archangel of Thune which is #3 on the rankings list. It was uncommon missing from the pack.

Does this mean that the current draft AI determines what colour it wants to be in first before the draft even starts? Or does it mean that the foils aren't ranked by the AI?

Ok, after testing and analyzing the code, i found a big logic error in the draft pick function. The AI was picking a random card from all cards that were at least as good as the last common in the pack.

Thanks gecktrix.

[lujo]

On that note, do you know if there's something going on with timeshifted cards from the Time Spiral set? The AI will pass them around no matter what they are, and that leads to unreasonably spectacular 9-15th picks for the human.

What's the AI stance on those and why?

-----

Oh, and COULD we get "replacements" for not-yet-implemented cards in boosters which contain them? A promo card, a "fabulous treasure" card which can't be played but can be sold in the shop for a bunch, anything really. I know it's more of a feature request, but still...

[CyanKnight]

AI does not favor Red and Black mana for paying the cumulative upkeep of Balduvian Fallen

[gecktrix]

Ok, after testing and analyzing the code, i found a big logic error in the draft pick function. The AI was picking a random card from all cards that were at least as good as the last common in the pack.

Thanks gecktrix.

I appreciate that you looked into that. Limited play is my favorite part of magic so anything I can do to improve it I will.

I do have a couple of suggestions that may or may not be possible to implement.

I'd like to see that when the AI has to pick a card, if the potential picks are relatively close to each other in the rankings I'd like there to be a chance that the AI will pick the slightly weaker card. Obviously not in the case if a verifiable bomb so if there's a card in the top 30 or whatever it will pick it up.

I'd also like to see that when it opens a bomb in a colour it is not in, there would be a low chance that the AI would consider splashing it or jumping into that colour and drop their weaker colour (the check for changing colours this should only happen in pack 2 and ideally if the AI can see that one of his colours is particularly weak). This will also apply for gold cards. For example, if the AI opens Garruk, Apex Predator pack three and it is already in green or black, it should highly splashing it. Of course it should not consider splashing a spell with double coloured mana costs like Anger of the Gods.

Lastly, i support lujo's posts regarding draft archetype support. I do not know if it is even possible, but the drafting thread I linked earlier does have interesting ideas regarding it. I have a lot more to add but now I feel like im rambling so this is all for now.

[lujo]

^ I think it's a set by set thing, and I'll have to finish restructuring the FE pick order list and try it out with the current rules to see exactly what's needed, and then we can start figuring out how to do it exactly. Making the array for "picks" have more fields for individual cards is probably a good idea in general, let's just be certain on what attributes should a card in a pack and in a set have for the AI to keep track of, juggle and evaluate before we start applying anything.

For example, if we know what the "bombs" are, we can tag them as such, etc. etc.

[lujo]

On a different note:

The AI blocks first, then uses Spore Cloud. This causes it's cretures to suffer the same effect as the attackers, which is the wrong way of using this card (it's the same as Tangle, only older so it has a stupid wording). It also tends to use combat pumps and abilities, and keep mana open for them, even though it's going to cast Spore Cloud.

Could it be made that the if the AI is going to use Spore Cloud it knows this before blocking and not block?

EDIT: The AI is actually losing game due to this, because it blocks with all of it's guys, while the attacker only sent it's strongest guys. So next turn the AI who misused Spore Cloud gets killed by the little guys XD

[gecktrix]

^ I think it's a set by set thing, and I'll have to finish restructuring the FE pick order list and try it out with the current rules to see exactly what's needed, and then we can start figuring out how to do it exactly. Making the array for "picks" have more fields for individual cards is probably a good idea in general, let's just be certain on what attributes should a card in a pack and in a set have for the AI to keep track of, juggle and evaluate before we start applying anything.

For example, if we know what the "bombs" are, we can tag them as such, etc. etc.

Yes it's definately a set by set thing. Personally, I feel that it would be more benefitial as a whole to have triple MMA, triple INN, or something more modern worked and tested on first for draftable archetypes. I do understand the Fallen Empires is one of your personal faves, and no offence, but the other sets are more realvent sets in today's world, plus they were generally a great set to draft with draft archtypes purposely built into the set. I'm definately still all in for improving older sets too.

I'd like to add with draft archetypes, we also would not want the AI to always draft it. For example, let's say that in triple innistrad, AI opens pack 1 and opens no bombs, but 2 good first pickable cards similar in ranking + Spider Spawning. (i know the spider spawning deck wouldn't be played optimally by the AI but this is just an example). Spider Spawning would be flagged as a lynchpin of an archetype that can be drafted around. The AI would do a check, and let's say 35% of the time it would go for good card #1, 35% go for good card #2, and 30% go for Spider Spawning, and starts to draft a deck around it. If the lynchpin card does not get drafted, it moves to the next AI, and again another check is run, but with progressively lower %s for the AI to jump into that archetype. By pick 4 or 5, it no longer becomes a factor, and the AI will no longer run checks to draft that archetype.

Just a dream.

[lujo]

If I'd chosen by personal prefference, I'd have chosen 3 X Oddisey, I picked FE for the following reasons:

1) I know the set very well, so I can't really mess up too hard
2) It's small, so I'm juggling 102 cards instead of 300 for the first try. So it's much easier to tell if whatever change you make is working and how it's affecting the whole thing.
3) It's fairly complex, so if I can handle it I can handle something larger but more straightforward
4) Quest mode sets you up with random drafts, and if you want to choose your next draft you have to play and win whatever comes up. So they're all equal in terms of relevance in Forge, it's just a matter of when I get to one
5) Older sets have the worst drafting AI due to the bestiaire lists being inadequate and the highest number of easily fixable bugs, so they need the most help.
6) Playtesting it is easier on the eye due to art for me. If I were the kind of "only work on what I'm interested" kind of guy, Invasion would probably be the last set I'd touch, followed closely by Mirrodin. However, I have the best intention to give all of them equal treatment. I just decided to start at the beggining.
7) FE, other early sets, Oddisey block, Time Spiral Block and Masques block first, as far as I'm concerned. Those were the interesting and weird ones. Of everything with the post-mirrodin frames, I think only Lorwyn, Innistrad and Kamigawa are really intersting, as for relevance - couldn't give a hoot. If I wanted to draft something recent, I'd be playing cardboard or MtGO. Not that I'd mind it working well in draft, but I'd prefer getting there from the other end.
And really, getting to draft MMA on Forge is meh, you don't need forge for that XD.


As for advanced archetype shennanigans, really, I'm currently only capable of making a big ole pick list. So Imma do that first. Then we see about more than that. And since I'm dealing with a set I actually know about, we can then see if stuff works easier

[schizophoenix]

When I have Underworld Dreams and the AI has Farsight Mask in play, the first time it draws a card, it chooses to draw cards until it dies from damage.

[lujo]

Guys, how can I locally disable the "Unusable by AI" tag so that I can test the drafting process better?

[ctleatherbear]

I apologize if this is redundant but season of the witch is not in the deckbuilder list.

gatherer dot wizards dot com/Pages/Card/Details dot aspx?multiverseid=1741