Requests for AI improvements

[HarlequinCasts]

Currently its AI base value is set to 40 (aka lowest priority for enchantment), so let's try to set it to 100 to see if it does anything...

Sadly I've seen no increase in the AI's desire to prevent itself from losing to my beloved Battle of Wits decks.

Here you can see it passed on using a Counterspell and then passed on using Disenchant before my next turn win.

[Korath]

Unsurprising. The exe doesn't use ai_base_value for anything except choosing which cards to keep during sealed deck construction; it's not wired into the AI's evaluation function at all. It's sometimes used in C for tutoring, too, but not much else.

[BAgate]

A card that has come up before, but I just had the AI cast an Arcbound Ravager and promptly eat the two artifact lands he had. Not a horrible move in and of itself, but I had a The Tabernacle at Pendrell Vale out, so he was dooming it to death. Is there any way to make the AI consider future upkeeps in its decisions?

[stassy]

AI keep choosing taking damage to death against Mogis, God of Slaughter even if it control creatures.

[HarlequinCasts]

A card that has come up before, but I just had the AI cast an Arcbound Ravager and promptly eat the two artifact lands he had. Not a horrible move in and of itself, but I had a The Tabernacle at Pendrell Vale out, so he was dooming it to death. Is there any way to make the AI consider future upkeeps in its decisions?

I don't think so, and even so Korath has mentioned before that the AI is just completely irresponsible when it comes to cards like Arcbound Ravager, Claws of Grix, Zuran Orb or anything that will let it sac permanents for a temporary gain. So the problem with that card goes pretty deep.

Also I'm pretty sure the AI only really maximizes its board presence on the current turn and can't really consider anything about future turns (hence always attacking even with a lethal counter attack on board).

[HarlequinCasts]

Just curious, do we have any idea why sometimes an AI just chooses to do nothing? Specifically what I mean is, I sometimes see an AI that will play no lands or spells, and then discard a land at the end of turn. Sometimes they may even play a Sapphire Mox or something, but will still discard a land.

[BAgate]

When does it do that? The only times I notice it are when it thinks it is about to lose and that was a deliberate attempt by Mok I believe to make games go faster. Have you seen it do that in the middle of competitive games?

[HarlequinCasts]

When does it do that? The only times I notice it are when it thinks it is about to lose and that was a deliberate attempt by Mok I believe to make games go faster. Have you seen it do that in the middle of competitive games?

Usually its all or nothing. As in, right from the start it will make (almost) no plays. I will often take a competitive deck through a gauntlet of 30 or 50 rounds so I occasionally see this behavior.

If maybe this is not as understood as I had hoped, Ill try investigating with replays next time. But as I said, sometimes it will literally discard land on the first turn, playing none, and go downhill from there.

[Nexhro]

The AI always casts Terminus for its miracle cost when it draws it, even when the player has no creatures on the battlefield and the AI does.
Would it be possible to make the AI evaluate its board position when deciding whether to cast Terminus?

[HarlequinCasts]

On the subject of my favorite deck - Battle of Wits

Something changed with the AI this patch! I'm now seeing the AI actively counter and destroy Battle of Wits, even with what are normally considered "better" targets. I've seen it discard its hand to Null Broach to counter it, and also I saw it kill a Battle of Wits using an Acidic Slime even though I had a Wurmcoil Engine. It even Recoil ed it to my hand to avoid losing!

What's new with the AI this patch? I'm actually losing Battle of Wits games due to the AI playing correctly for the first time ever. I like what I see!!!

[Korath]

It and a bunch of other alternate-win-condition cards got specific hinting in commit f6a05bb:

Code: Select all

$ git log -1 --grep "Battle of Wits"
commit f6a05bbe33b1438d1d24389e1d7acd3f3b7f23ac
Author: Korath <dgk@Dirge.(none)>
Date:   Sun Apr 6 23:36:00 2014 -0400

    Attempt to teach AI about alternate-win-condition cards

    Particularly those that trigger at upkeep, which otherwise won't be considered
    at all on the preceding turn.

    Affects cards:
    Azor's Elocutors
    Barren Glory
    Battle of Wits
    Chance Encounter
    Epic Struggle (and mark ready)
    Felidar Sovereign
    Hellkite Tyrant
    Immortal Coil
    Mayael's Aria
    Maze's End (and can't activate if animated and summonsick)
    Mortal Combat
    Near-Death Experience
    Test of Endurance (and use C version)

Basically, if you have a Battle of Wits in play and at least 200 cards in your deck, the AI thinks it has -99 life (no matter the current phase or turn). So it might try to mill you, too, if that can get you below 200 with a single spell or activation.

One thing it won't do is try any harder to cast Time Walk or Eon Hub or such so that you don't get an upkeep for it to trigger in, and if cards like Stifle worked, it probably wouldn't cast it to counter the upkeep trigger either. Though the AI hardly needs more encouragement to cast Time Walk, I guess.

[HarlequinCasts]

It and a bunch of other alternate-win-condition cards got specific hinting in commit f6a05bb

Fantastic! After another day's playing I have yet to see a situation where the AI fails to play correctly with a Battle of Wits on the board. Including some seemingly clever plays. Just like you say, it does almost anything and everything to ensure it does not lose to it.

Much more fun now!

Since I'm always talking about it, if anyone is ever interested here is my Modern legal Battle of Wits Esper deck.

[HarlequinCasts]

One thing it won't do is try any harder to cast Time Walk or Eon Hub or such so that you don't get an upkeep for it to trigger in, and if cards like Stifle worked, it probably wouldn't cast it to counter the upkeep trigger either. Though the AI hardly needs more encouragement to cast Time Walk, I guess.

Interesting, I actually saw a scenario where the AI had Voltaic Key and Time Vault in play (the AI has never been good about actually taking unlimited turns). Despite having sort of randomly taken many extra turns, it failed to decide to take an extra turn when I had a Battle of Wits out, thus losing. I assume this is the same logic as Time Walk where it does not know to avoid losing by denying me a turn?

[Nexhro]

Would be cool if the AI could stop targetting itself and/or creatures controlled by it when it gets to choose targets for Karplusan Minotaur 's cumulative upkeep.

[Korath]

Is it your Karplusan Minotaur or the AI's? If yours, does the AI pick targets any better if it controls the minotaur instead?