Collectible Card Game Headquarters

AI activates Chandra, Pyromaster +0 ability, exiling Scavenging Ooze. AI has Misty Rainforest and one other untapped land. AI does not crack the fetch to cast Scavenging Ooze, presumably because it does not see any cards in its hand that it needs to crack the fetch for (but the Scavenging Ooze is not in its hand).

opponent repeatedly plays Fossil Find over and over (this being the only card in their graveyard after cast, as in 1 in hand 1 in gy, they do this until their mana is depleted)

AI (quest opponent Elrond) had Krond the Dawn-Clad in play enchanted with an aura and it attacked me - and exiled itself. First time i played against this card, don't know if the AI never uses it correctly (i had no creatures in play), but this clearly isn't how it is supposed to play this card.

| Open

Archfiend of Depravity

from FRF makes you sacrifice all your creatures when it should be all but two instead. thanks.

http://magiccards.info/frf/en/62.html

Just wondering, but the AI cannot really play any constructed decks, nor combo (like storm), it just plays on draft / sealed (e, one card at a time, no overall strategy for a deck), correct?

jmlima wrote:Just wondering, but the AI cannot really play any constructed decks, nor combo (like storm), it just plays on draft / sealed (e, one card at a time, no overall strategy for a deck), correct?

"one card at a time, no overall strategy for a deck" is correct. That said, the AI can still perform decently with most Standard/Block decks if the deckbuilder is aware of the AI limits (ie. replace Chandra, Pyromaster with Outpost Siege, Abzan Advantage with Erase, etc.). Just think of it as a simple player that tries to maintain tempo by playing its most expensive card in hand unless it has the right answer (Disdainful Stroke, Lightning Strike, etc.) for the situation.

The AI had Renegade Doppleganger in play, then cast Bonehoard. Once Bonehoard resolved, it cloned the Germ token with the Renegade Doppleganger, causing it to die immediately. The AI should not clone */0 creatures unless the clone has some sort of toughness buff.

Hi,
First time I post in this thread, always accepted most of the AI stuff.
I did post some AI remarks when I launched the Mirrodin world, not sure if these have been noticed:

| Open

Mirrodin has 1147 cards, almost 20% can not be played properly by the AI, these cards are not in the opponents decks. You won’t be battling against:
- Sunburst
- Imprint
- Sacrifice artifact cards
- The plethora of “machine” combo’s that this world has to offer (Blasting Station, Myr Galvanizer combo, Triskelion, Clock of Omens, Staff of Domination, Mycosynth Lattice, Eggs, etc.)
- several of my favourite cards: Venser, the Sojourner, Mimic Vat, Proteus Staff, Death Cloud, Isochron Scepter
- and yes, i tried them all in dedicated decks, and they let all me down Some simple AI code however would allow dedicated decks with these cards, e.g.: AI uses Proteus staff to target their own creatures; Mimic vat: don’t activate it when nothing is imprinted; don't play Isochron Scepter when no valid cards in hand.
The good news is: it is up to you to discover and build around all these cards!

I used 4 cards that are deemed not playable by the AI according to Forge. They work however sufficiently well in the decks i crafted for them. These cards are: Birthing Pod, Dispatch, Magma Giant, Endless Whispers

Karn Liberated is listed as playable by the AI, but in my opinion he is not. Are planeswalkers coded that they only use the + ability and the ultimate? Too often the wrong ability is chosen resulting in a loss instead of a certain victory. Tezzeret does the same, but remains strong enough to earn his spot in a few decks.

My main concern with battling the AI is that it always tends to overcommit. I’m a control player, so I never go anywhere without my wraths. And wrathing for good value is too easy.

Has there been attention for this behavior? Could the AI be improved? Better analyzing the board state and some probabilities won’t do it if the wrath in my hand is not one of the scenario’s.

I first thought of the following:
- Second main phase: if the AI attacked and I lost more than 20% of my life total as a result: don’t play a new permanent, things are going fine.

Unfortunately there are other scenario's:
- I’m mana screwed/flooded
- we are racing
Then AI should definitely commit as much to the board as possible.

Hence not overcommitting should be something the AI only does based on the information it has:
- if in one of the previous games of the match I did play a wrath than the Ai should not overcommit (see playrule above)
Me playing a wrath could be identified by classifying all cards in Forge and check if I played such a card, or by noticing that I played a spell that killed > x creatures.

I know very little about AI design and how it’s being handled in Forge, so this post is mostly done out of curiousity (and I couldn’t find more about this in this thread).

This also touches sideboarding by AI, something I could also only find scarce comments of in the threads but undoubtly has been considered and discussed.

The information from the previous game log could be easily translated to some default sideboard options. It would be great fun and a great improvement to have the tools to define these for all decks.
A simple approach:
- the AI always takes out the same 4 cards and has 3 scenario’s to choose which 4 new cards come in
More complicated:
- it can be defined that any card goes out and any card goes in
I could see some standard rules you could pick (or create) as a Forge deck builder and with which you can assocociate some main deck & sideboard cards:
- if in one of the previous games of the match the AI has been targeted > 4 times then.. (4x Leyline of Sanctity to replace card x)
- if in one of the previous games the humanplayer controlled/played > 3 nonbasic lands then.. (Blood Moon)
- if in one of the previous games the humanplayer played < 1 creature then.. (take removal out)
- if in one of the previous games the humanplayer played > 3 enchantments/artifacts then.. (Disenchant)
etc.

serrasmurf wrote:I did post some AI remarks when I launched the Mirrodin world, not sure if these have been noticed:

The AI is usually really really bad at using planeswalkers since it evaluates each choice in isolation AND order. The rare cases where it does well like Elspeth, Sun's Champion is due to sheer luck as it will not use the +1 in Main1 usually, allowing it to evaluate the -3 so naturally it will fall back to the +1 in Main2. I pretty much hacked the + abilities of Domri Rade and Ugin, the Spirit Dragon to even allow the AI to evaluate the minus ones...

For the other block/set/etc. specific mechanics and interactions, the AI is generally unaware of them unless someone scripted them in as AiLogic or some hard-code.

serrasmurf wrote:My main concern with battling the AI is that it always tends to overcommit. I’m a control player, so I never go anywhere without my wraths. And wrathing for good value is too easy.

This goes back to the problem of the AI not having any "strategy" per-se or knowledge of its own deck, much less the opponent or metagame. As I answered in an earlier deck, the AI is coded to evaluate each priority step in isolation except for specifically coded cases.

It is generally better for the AI to over-commit to race and force the opponent to make sub-optimal plays in response or die. It may lose to that sweeper in hand but the likelihood of facing a sweeper and/or a control deck that can take advantage is usually less than the easy wins it can get. Of course if you prefer to play such decks, it would make sense to tailor the AI deck with resilient threats (ie. Ashcloud Phoenix, Whisperwood Elemental) .

serrasmurf wrote:This also touches sideboarding by AI, something I could also only find scarce comments of in the threads but undoubtly has been considered and discussed.

This is probably too far out considering the points above... Personally, I construct 75 card decks and play against pre-sideboarded AI opponents for some semblance of fairness.

excessum wrote: Of course if you prefer to play such decks, it would make sense to tailor the AI deck with resilient threats (ie. Ashcloud Phoenix, Whisperwood Elemental) .

... Personally, I construct 75 card decks and play against pre-sideboarded AI opponents for some semblance of fairness.

Thanks for your response! I mostly play Quest, so optimizing that experience (and AI resistance) is my interest, hence the ideas.

Is using information from previous games to create an "AI strategy" feasible or would make that things too complicated?

serrasmurf wrote:Thanks for your response! I mostly play Quest, so optimizing that experience (and AI resistance) is my interest, hence the ideas.

Is using information from previous games to create an "AI strategy" feasible or would make that things too complicated?

I doubt that this is going to be feasible since most quest decks are themed and would require manual tinkering to fit a 15 card sideboard in. The AI is definitely not able to identify the "threat" card(s) in its opponent deck like AEther Vial, Pyromancer Ascension, Show and Tell, etc. so a generic answer is probably more harmful than good. Choosing which cards to board out is another issue...

excessum wrote:I doubt that this is going to be feasible since most quest decks are themed and would require manual tinkering to fit a 15 card sideboard in. The AI is definitely not able to identify the "threat" card(s) in its opponent deck like AEther Vial, Pyromancer Ascension, Show and Tell, etc. so a generic answer is probably more harmful than good. Choosing which cards to board out is another issue...

You could very well be right that the gain of these features would be limited or even counterproductive. So let’s do a quest-metagame-sideboard analysis. Feel free to add/ criticize.

Let’s try if we can limit ourselves to the basic 3 archetypes: aggro, control & combo.

AI-Quest-opponents
90% of these decks can be considered as aggro, 5% combo and 5% control. The numbers change a bit when you reach the higher levels.
This makes sense as the AI can handle aggro best. Aggro requires limited strategy, just dump creatures on the board, synergy often does the trick as most of these decks are very lineair.
Combo is complicate for the AI to pull off because
- it can't play the cards properly (marked as such in Forge)
- it doesn’t recognize 2 cards being a combo
- it can’t time it properly
- it can’t protect it properly
Control is also tough because the AI:
- cannot distinguish between unharmful threats (leave be) and dangerous threats (use removal)
- cannot cherish it’s scarce win conditions
- Is disadvantaged in the long game as it gives the humanplayer more possibilities to outplay the AI

Humanplayer
I have no data on the type of decks everybody plays (would be interesting survey), here an estimation based on forum discussions.
In quest you mostly start with a sealed pool, more or less customized in the direction you want to take. The first deck often has several creatures, some control elements and some bombs that win you most of your games. I’d qualify these as aggro-control. From there you develop your deck to:
- More aggressive & burn - ?%
- Solid (lineair) creature deck -?%
- Control deck -?%
- aggro-control-%
Playing an Ensnaring Bridge deck with a lot of small creatures is an aggro-control deck to me. Playing ensnaring bridge with more control elements and a few win conditions is a conrol deck.
Combo is most of the time only feasible as a (small) plan B, as it is very hard to find enough copies of your combo pieces and the support cards needed to make it even a little bit consistent.
And there are also several players who preserve the nature of the sealed deck by throwing all cool cards they can find in the deck. Vorthos players do the same or try to build lineair aggro decks as they buy all the zombies they can get their hands on.

The humanplayer has the opportunity to adapt to the "metagame". As 90% is aggro it means for instance main decking creature removal and fill your sideboard with counterspells and enchantment removal.

Now we have some idea of the human-oppoents. What could the 90% aggro AI decks do?
Aggro versus aggro:
Here is not much room for strategy. It is mostly more a matter of how the decks play out. Goblins win from big creatures. Mid-range wins from goblins. Big creatures win from midrange.
Sideboarding more removal will be counterproductive.

Aggro versus pile-of-cards (sealed pool)
Not much to do here too. As there is no prevalent theme within the humanplayer deck there is also nothing to adapt to, whether this is style of play or sideboarding.

Aggro versus control:
Here I do see room for improvement. Not overcommitting would be better. Adapting playstyle based on the information from previous games would be an interesting experiment. As the aggro decks are very lineair there is limited space for sideboarding though.

Sideboarding for AI seems not be very effective. But it will be great fun, so that’s worth something too:
- Building AI decks will be more interesting with sideboards and sideboard rules
- Playing against AI decks will be more exciting as they might surprise you with nasty sideboard cards or a complete transformative sideboard (change 15 cards and let the deck play different)

serrasmurf wrote:AI-Quest-opponents
90% of these decks can be considered as aggro, 5% combo and 5% control. The numbers change a bit when you reach the higher levels.

The distinction of Aggro/Combo/Control is a bit misleading here. If you include Midrange (and most magic writers agree to include Midrange as a 4th archtype), the numbers are more like 75% Midrange, 10% Control, 10% Combo, 5% Aggro.

See this link for quest deck archtypes: viewtopic.php?f=48&t=11538

But i doubt that the AI can identify the human's archtype reliably anyway.

excessum wrote:Choosing which cards to board out is another issue...

This is the main problem of AI sideboarding.

The only possibility that i can imagine is manually marking cards in the decklist to board out (maybe even as an ordered list). And manually adding sideboard hints to cards, so the AI will board them in under given circumstances (like adding "AISideboardHint: Card.Blue GE 20" to Red Elemental Blast).

AI just used Sarkhan, the Dragonspeaker's +1 ability and made him a legendary creature then played another Sarkhan, the Dragonspeaker that ofcourse died.

I guess there is a check missing there.

The AI is still incapable of using Knight of the Reliquary , it just keeps recycling lands Ad-Nauseam never doing anything with the 15-15 it generates, it does not even use it to block when its the only thing that can prevent it from losing.

The AI box should note that this is one of the cards the AI does not know how to play with.