It is currently 20 Jul 2018, 14:49
   
Text Size

AI Improvement suggestions

MicroProse's Shandalar Campaign Game, now with new cards & a new look!

Moderators: drool66, stassy, BAgate, gmzombie, Aswan jaguar, CCGHQ Admins

AI Improvement suggestions

Postby lujo » 04 Nov 2016, 01:06

I've taken it upon myself to build a nice exclude list for my deckpack, and general purpose use. I don't, however, feel like I should recommend that folks exclude things from the pool without at least giving them a look to check for bugs. So it's something I'm doing now, fiddling around with cards I don't intend to use much to check them for bugs. And I do a lot of playing around with a lot of cards because I make and tweak AI decks, anyway. And Korath has said that he isn't looking to do any individual card AI improvements on any schedule but his own. But I still get to see all kinds of stuff and not all of it is obvious or intended and it might just happen that something is easy to tweak.

The point of this thread is to let me separate the many "AI routinely mishandles" from actual plain-bugged reports. Feel free to join with stuff you've noticed if you can describe it well enough. Most likely nothing's going to happen, but wth, if you're making decks and you know something is being mishandled - I'd appreciate all the time you saved me by informing me not to use something that's routinely mishandled.

These are all at 5000+ thinking time, HS2 , Wizard "resume game". I haven't looked up any code for any of this, but I have been patient and tried my best to be as thorough as possible while observing and testing in debug mode.

#1 AI Improvment - Alpha Wards | Open
( Black Ward , Blue Ward , Red Ward , Green Ward , White Ward )

The AI tends to only cast these on a dude if you have any permanent of the appropriate color. The problem is that these aren't "pick a color" wards, and waiting for me to have a permanent is bad for the AI. If one of these is in the AI deck, it doesn't have a reason to wait.

This is easies to grasp if you take Red Ward. If the AI has Red Ward maindecked, then it wants it on one of it's dudes regadless of what I have on board, because it's got something going on with that card. I may not ever even play a red permanent or have a red permanent in my deck at all. If the AI sideboards this vs. red, which is more likely, waiting for me to play a red permanent is going to get something killed with Lightning Bolt while this stays in hand. Or that permanent might be Seal of Fire or Flametongue Kavu.

It feels like these things are using Scarab AI, but they aren't the same at all, and you can make decks with these maindecked (Pestilence / Pyrohemia for example) distinct from the ones which you would with scarabs (They're more Painter's Servant / Mind Bend y sorts). They're also not pick-and-choose wards so that their wielder has to wait to see what they're playing against.

AI seems to be appropriately eager to play Pledge of Loyalty with nothing on my side of the board, maybe these should be using whatever it is using?


#2 AI Improvement - Artifact Ward | Open
I think this thing has some serious issue with it - I think it only gets played if there's an artifact creature on board or something. It, like #1, might be a remnant of the original AI "thinking of" protection (and landwalk) primarily as a means of evasion (as opposed to the more insidious hexproof+evasion+semiindistructible that it is).

The AI seems extremely reluctant to play this, and I've had it staring at Flamecast Wheel and Aeolipile and refusing to play it. It's probably the same AI as for the alpha wards, except those seem to look for permanents while this probably just looks for creatures.

Still, same advice - it's not multiple choice and the AI is (see above) no the best at assessing threats, so if this is in an AI deck the right call is to play it on something ASAP. Either the deck needs something to be imune to artifacts for it's own ends, or it sideboarded.


#3 AI Improvement - Flickering Ward | Open
Has several problems. It's specific in that it can self bounce. Giving it to the AI in this state is a bad idea, it seems.

When it's cast it seems to count the number of creatures of each color that the opponent controls and chooses the most frequent color. If the AI played a white 1/1 and you played a white 1/1, it's going to be on his 1/1 and give him protection from white. However, if then next turn you cast Dark Ritual and throw down a Juzam Djinn, the AI won't bounce it to give its guy the much more adequate protection black unless you cast another black creature at some point.

The bouncing effect seems to be locked down tight (possibly to prevent the AI from bouncing it too much?). The effect is that as long as it's stuck in the "right" color (meaning the most frequent color among creatures on the other side) the AI seemingly won't consider bouncing it even to move it from a 1/1 to it's biggest/most important creature.

This is, well, unsuitable AI for this card. I'm not sure what the suitable AI would be, but I wouldn't mind testing it with other parameters, like total power or total power+toughness, or total untapped power and toughness on the other side. And definitelly more willingness to put it on it's bigges stuff, as with its cheap cost it's bound to get stuck on a weenie.

The other problem is that it doesn't seem to have any in-built AI for cooperating with Heroic. You could make a Heroic deck with this as a centerpiece easily, but the AI doesn't recognize the potential of it. I've had the AI with a lot of white mana, Flickering Ward on an Advance Scout doing nothing and a Satyr Hoplite sitting there also doing nothing turn after turn.

The one good thing, apart from it passing the strictly technical "It's not bugged" exam is that the AI will attempt to bounce it when the creature it is on is about to die, props for that.


#4 AI Improvemen - Cho-Manno's Blessing (Wrong Phase) | Open
It's cast during the AI's EoT instead of opponents. This makes it difficult to tell much else about what the AI's doing with it. I think I could also just report this as a honest-to-god bug.


#5 AI Improvement - Floating Shield (Extreme reluctance) | Open
The AI seems extremely reluctant to play this if it has only one creature on the board and in hand. I've had up to 6 red creatures to its 1 white one and it still didn't play it. As soon as it is given a 2nd creature (into play or into hand) it does play the aura.

I think it's just bugged or overly biased to only be played if the AI has more than 1 creature somehow. I can't tell how well it's being handled past that point before that is cleared up.


#6 AI Improvement - Madness Wards / Madness in general | Open
Strength of Isolation and Strength of Lunacy will get played for their regular cost even if there is a way for them to be played for their madness cost and even if there is a discard outlet available. This might be an "alternate cost" thing, but I'm still noting it.

What is worse, the AI will generally play as if it has no way to distinguish between Madness cards and regular cards in hand when self-discarding. So you will see things like the AI holding Strength of Isolation in it's hand and discarding 3-4 other cards for no obvious gain before it discards it, triggers Madness and pays for it.

It does occasionally make the correct move right away, but that seems like an incident. Even when cards are discarded by enemy effects, it seems to be as likely to discard lands as it is Strength of Lunacy or even Dodecapod.

It's safe to say not a single card with Madness can safely been given to the AI in the capacity of a Madness card, currently. The obvious improvement, for a start, would be for the AI to check for madness cards in hand whenever it is prompted to discard for any reason, then check whether it has mana to pay for Madness and whether there are targets if those are needed, etc, etc.


#7 AI Improvement - Patrol Hound / Discard outlets | Open
While trying out the madness wards I happened to observe a few discard outlets, namely a few spellshapers and cards like Patrol Hound. The unbounded discard guys like the Patrol Hound seem to be absolutely unsafe for the AI.

Firstly, the AI is too eager to discard random cards, at odd times and frequently. Patrol Hound routinely got given first strike at the end of it's owner's turn, after blockers even if unblocked, at the end of the enemy turn, and possibly at other times.

The lack of a "prompted to look for Madness cards in hand when prompted to discard" meant it would routinely burn through any number of cards before hitting the one which made sense to discard.

Also, none of these activations make any sense unless there is a discard friendly card in hand as an effect. Patrol Hound most likely doesn't need first strike unless it's in combat and fighting something, so he most probably should not be discarding anything "at specified times".

Discard outlets confused a lot of people because their proper use is quite complex and their great use it very context sensitive. It's also outlet-specific, Wild Mongrel and Patrol Hound are rather different cards. I'd say it's impossible to have universal AI instructions that would make the AI use them perfectly in any situation, but they can be put together to be used correctly in some contexts, those contexts noted and then usable in deckbuilding.

As it stands a safe bet is that "Discard a card" sould be preemptively read by deckbuilders as "Discard a card at random at any possible time" when considering making decks with discard outlets (ones similar to Patrol Hound especially) in them.

As a general improvement, at the very least, discarding which was meant to be encouraged during the enemy EoT step, should not be happening during the player EoT step. This seems to be a common misplay.


#8 Deckbuilder Heads Up - Mother of Runes | Open
While testing Mother of Runes I ran into a number of instances of odd behavior. If there isn't something definitely wrong with her particular AI, she is prone to randomly be affected by something.

Namely, I've had her be untapped and not save the only other creature the AI had on board from a Swords to Plowshares for no discernible reason. Next turn, she did protect the same creature under the same circumstances. When she'll attempt to save something (including herself) and when she won't is hard to determine and she doesn't seem to be primed to always attempt to save things.

Another time the AI reacted to a Mind Burst (random example) by tapping her to give herself protection. She gave herself protection from red, which seemed to be a random choice (nothing on the board or in anyone's hand was red). She didn't do this next turn when the AI got hit with Mind Burst again.

So she's unreliable. And from what I've seen the AI is somewhat prone to using her as a source of evasion. She's rather more likely to be put into decks in order to prevent removal from happening so this might not be what the deckbuilder may like. She's rather more likely to attempt this with Vigilant creatures it seems.

As for improving the AI, I'm not sure why she would react to spells/abilities which have her owner be the only target. Brigid, Hero of Kinsbaile , Lavalanche , Ember Gale , Heart of Bogardan , Chandra's Fury , Aggravate and Flame Wave are the only cards that come up with "target player" and "controls" that her effect can do something about, while there are 363 cards in Shandalar which say "target player". Not all those would be cast at the opponent, but if she's going to be fiddly she might as well not potentially tap herself in response to a lot of nonsensical spells.

Also, I'm not sure what it is with the random "wouldn't tap to save dudes targeted by removal", if she is untapped. That's kinda her only job.

And I haven't even tried to check whether the AI is any good at playing against her and how the AI reacts to her shennanigans.


#9 Deckbuilder Heads Up - Eight-and-a-Half-Tails | Open
Don't give this card to the AI, as it is severely mishandled. The AI exibits the same behavior as with Mother of Runes, meaning it sometimes will and sometimes won't save things from removal, which is more or less what this card is for at a basic level.

The pattern seems to be that casting removal in the AI's turn (even at Eigth-and-a-Half-Tails) is likely to not prompt a reaction from the AI. Upkeep seems to stump it, as does end of it's second main phase.

However, a specific, well, bug I suppose is that the AI is quite likely to change the color of it's lands into white and randomly give them the protection for no apparent reason. It seems to do this a lot. I'm not sure if it only does this when it has mana to spare and is fooling around (sometimes the insane behavior comes out in those situations), or whether it would do this instead of doing sensible things.

Not fit to be used in decks I'm aftaraid. The land bussiness needs to go, for one thing. Only target lands if they are creatures and/or are targeted by removal?


#10 Deckbuilder Heads Up - Protection from color soaks trample up | Open
I wasn't aware of before, but I've just now seen Spectral Lynx block Spectral Force and then nothing happned. Because protection soaked up all the Trample damage. This is pretty messed up, and it's a good enough reason to give protection from color cards a wide berth when making decks for the AI as well as not having them in the pool for the player. Trample is one of the few ways around protection from color.


#11 Deckbuilding, difficulty and +1/-1 - Spirit of the Night | Open
The AI is quite eager to pump creatures like Breathstealer and Metropolis Sprite to death at lower difficulties. Breathstealer is used with Urborg Panther and Feral Shadow to bring out Spirit of the Night , and the AI is capable of this - if it doesn't first kill it's Breathstealer. And to a lesser degree if it doesn't sacrifice Urborg Panther agressively, which it is also eager to but that's not really a misplay.

The important takeaway is that if you make decks with creatures which can repeatedly pump power while lowering toughness ( 25 creatures with just "+1/-1" , but there are other ones ) , on some difficulties your deck will work and on some difficulties it will fail spectacularly. Possibly at all difficulties if you happen to add something else which confuses the AI in the exact right way (I haven't been able to figure out what does that).

The improvement has been posted as a feature suggestion allready, just make the AI use the same AI capacity at every campaign difficulty. Or at least allow this to be toggleable in the .ini. This tendency affects things beyond making one creature kill itself, as seen from the Spirit of the Night thing.


#12 Heads-up for Invasion Master cycle | Open
I was testing Stormscape Master and Thornscape Master to see what they do with their protection granting ability.

Stormscape Master just kept using her drain ability at the end of her owners turn, rather than opponents, and never once did anything else. The mistiming should be reported and fixed I suppose, no point in her doing this at that particular time.

Thornscape Master stayed open and did things in response and did the right thing often enough, but he also had screwups like targeting himself instead of the creature about to be removed, and giving himself protection from the color of a creature attacking his owner, but did this before he blocked rather than after. Also, I've seen the AI block with another creature while he was untapped, not give that creature appropriate protection, and then not finish the enemy off with the red ability.

All in all, these two seem like a liability if given to the AI, with a good start at making them better being proper phases flagged as when to activate their abilities. I guess.


#13 AI Improvement - Ward Sliver | Open
Ward Sliver appears to not have it's color picking AI set straight. I had the AI cast one with 2 black creatures on my side of the board, and the AI chose black which seemd ok and consistent with various similar cards. Then I gave myself three green creatures to see if it changes to green with the next Ward Sliver. I gave it a Ward Sliver to cast, and it chose white, even though there were no white permanents on board. Then the next one also chose white, and the next one also chose white.

No idea what is going on there. Unless adding cards to opponents hand in debug mode is what's causing strange behavior, and it might be, who knows, Ward Sliver needs it's color picking scheme looked at. Its role is most commonly that of the finisher and the protection is most likely be intended to give evasion, so going for the color with the highest number of enemy creatures is probably the best choice.

Someone who's really into slivers ought to make a simple and basic sliver deck with 4 ward slivers as evasion granting finishers and play a bunch of it versus monocolor decks (at Wizard 5000+ thinking time) , see if the AI makes bizzare color choices.


Here I really have to wonder whether it's the debug mode causing #13 and #14:

#14 AI Improvement needed for Voice of All | Open
It seems to suffer from the same problem as Ward Sliver. I had the AI play 4 Voice of All in a row against a blue board on my side. It choose white every time. I had Alexi, Zephyr Mage, so I bounced them back. Four Voice of All against 6 blue guys, chose white all 4 of them. Then again and again, and every time the same thing.

Now, they're not Ward Sliver so they're not looking to give mass-evasion vs. as many enemy dudes as possible, and they fly so if evasion is the primary motivaition they should probably be looking for the most frequent color vs. enemy flyers, but I'm really not sure what is up with consistently picking white.

The only explanation I have is if it picks based on what's in the enemy deck/graveyard.


#16 Defender of Law / Flash | Open
Not sure if it's a matter of properly flagging individual cards, but the AI played Defender of law in combat, before blockers, on it's own turn. It always looks strange when this happens and I'm not sure why it is happening. It looks like a common occurence not just with cards but also with abilities like the one on Homarid Spawning Bed.

It's probably meant to create a surprise blocker, but the AI doesn't distinguish who's turn/attack phase it is.


#18 Wall of Dust and the AI | Open
The AI seems to be strangely reluctant to attack into Wall of Dust. It has a creature which can get flying, but it doesn't even bother. It can band two guys to be able to kill the Wall of Dust if it blocks - doesn't even bother. Has five creatures against one Wall of Dust - won't attack. Default AI attack prefferrence settings.

I'm not sure if it's about Wall of Dust or just combat AI but is rather looks like it's overvaluing the Wall of Dust aftereffect. I'll try it out some more, I'm making a note of it to reming myself.


#19 (Merge #15 and #17) Pump Knights | Open
"Pump Knight" cards, XX, 2/1 (or 2/2) with protection and two activated abilities, one that pumps their power and one which gives them first strike or flying seem to be hazardous or at least to undependable to give to the AI.

#15 - I just saw the AI give first strike to an Order of Leitbur which was blocked by a creature with protection from white.

Also, I saw an unblocked Order of the White Shield get pumped only 3 times when it could've been pumped 8 times. The AI had nothing in hand and nothing but the two pump knights to do with it's mana.

#17 - I've observed Stromgald Crusader and White Shield Crusader a bit. The AI seemed to pump Stromgald Crusader before combat and before blockers, sometimes tying up important mana. The obvious thing to do might be to limit when it considers pumping, as the current state makes it a loose cannon.

White Shield Crusader simply wasn't being activated. The AI wouldn't give it flying in order to attack, and it wouldn't give it flying in order to block fliers. The ability works, it's just that the AI was really reluctant to use it.

#19 - Stilmoon Cavalier seems VERY eager to suck up all available white and black mana the AI has. Before combat, too. It's not exactly mindless ability spam, as it doesn't seem to attempt to stack First Strike and Flying activations, but it looks strange and seems to causse odd behavior with other cards which have activated abilities. See #20.


#20 - Jareth, Leonine Titan and ability spam issues (I think I'm onto a bug, possibly a known one) | Open
I've had the AI sometimes make nothing but reasonable plays with Jareth, Leonine Titan. Mostly when there's not much on the board. Protected itself from removal, made itself unblockable, and the blocking pump seems to be working right.

But I've seen the AI be overeager to activate his protection ability and pick nonsense colors until it manages to "hit" the one that would make it unblockable. What seemed to set him off was Stillmoon Cavaleer - the spam of various activations from Stillmoon Cavalier (pre combat) prompted nonsensical activations from Jareth.

This was quite reminiscent of how Breathstealer and Metropolis Sprite seemed to work perfectly until activation spam from Ravenous Bloodseeker and a few madness pump cards (some targeting Breathstealer) caused the AI to confuse itself into pumping it's own dudes to death.

What I think is the problem is that activated abilities flagged to be made in response to things are too open-ended as to what exactly they are supposed to be activated in reaction TO. Stillmoon Cavalier activates one of his things, Jareth reacts to that, but since Jareth isn't in any danger the color he gets protections from gets picked at random. Then Stillmoon Cavalier hits it up again (or even in response to Jareth), and then Jareth also hits it up again, and they confuse themselves into creating a mess.

So, I suppose Jareth, Leonine Titan should activate his ability in response to spells or abilities which target him as opposed to anything at all? At the very least? Is this a known thing?


#21 Deckbuilder warning about cards with repeatable activated abilities derived from #20 | Open
Cards with repeatable activated abilities might be quite hazardous to give to the AI, judging from several of these notes. They have a somewhat high frequency of their abilities being activated as if the AI "can't tell who's turn it is", but they also seem to have loose response triggers for when the AI is prompted to use them.

Putting several different cards with repeatable activated abilities in the same deck could cause very bizzare looking plays which come about due to the AI prompting itself to react to it's own ability activations with other ability activations (sometimes at the oddest times due to choosing to activate things at the right phase but on the wrong player's turn).

Even putting just one such card in 4X copies could, in theory, cause two copies of it to trigger each other into nonsense and/or self-harm. Or so it seems. What I'd do when building decks meant to be used by the AI "for serious" is stay very far away from highlander or EDH-like singleton card-piles because that increases the chance of this sort of thing randomly happening and makes it much harder to catch "inevitable" instances of it while testing the deck.


#22 Infect isn't properly implemented for the AI / Phyrexian Crusader | Open
There is no proper AI for Infect, just like there is none for Wither, and the AI can only use cards with it sensibly by accident.

I've just had a Phyrexian Crusader not attack into a 0/7 wall and also not attack into a 1/3 blue blocker turn after turn. Several things are wrong/odd there:

a) The AI is generally capable (and depending on the circumstances and difficulty also eager) to attack with creatures that can't die. This means both attacking into walls with 0 power and also attacking with it's own creatures with 0 power. It's quite baffling to then see it not attack with a creature with Wither/Infect into a lone Wall of Ice.
b) It seems to have no awareness of Wither/Infect creatures being any different from vanilla. This makes sense beacuas it's the same way about Deathtouch (up to HS2) , but also Flanking and Bushido.
c) Phyrexian Crusader in particular , and Wither/Infect + Frist Strike in general, likely need their own AI on top of some general Wither/Infect AI. The AI (obviously) doesn't recognize that creatures without First Strike that fight the Crusader will have their power lowered by the ammount of his power when they deal damage. It's impossible to play this kind of card right otherwise.

In light of this, I'd say it's pretty safe to say Wither/Infect shouldn't be given to the AI. I'll still have to test it all, but the potential disconect between how it's supposed to be used and how it seems to be used makes me wonder why it was even implemented (unless it was meant to be player-only until further notice).

EDIT: It is only somewhat sensibly used with fight cards, as the AI seems to consistently choose to use Rabid Bite targeting its Infect creature to "debuff" an enemy creature in appropritely constructed situations. But it does that with Deathtouch too, even though it's unusable in combat in HS2


Upon further testing:

#23 AI can't play against Infect/Wither at all | Open
Makes sense, since there's no recognition of Wither/Infect in combat the AI makes attacking and blocking decisions as if those creatures were vanilla. The possible exception might be when it has many poison counters.

The result of this are predicatably horrible plays vs Phyrexian Crusader both on offense and defense. The AI attacks into it with 2/3 guys and blocks it with anything with toughness higher then 2. It's quite simmilar to how the AI simply can't handle Voracious Cobra - due to the exact same reason, up to HS2 Deathtouch wasn't properly recognized by the AI and you can imagine what it looks like if the AI is playing against Voracious Cobra but doesn't understand it has deathtouch. (I'm curious whether what Korath did with deathtouch for the next release made the Voracious Cobra situation any better). It's not much better against other creatures with Wither or Infect.

So the AI can't use Wither/Infect and it certainly can't play against it, which raises the question of what it is doing in the pool at all in this state. Don't give it to the AI and don't use it against the AI, it can't possibly function except by accident.

This isn't about "poison", I haven't gotten around to poison, but Infect isn't the same thing as poison and even if the AI can poison you to death with Infect creatures if enough go through and deal damage to you, the AI is likely playing them very badly, certainly nothing like they're supposed to be played, in a way confusing to anyone who knows how they are supposed to work and is certainly incapable of playing against them.


#24 AI Improvement for Card: Wall of Putrid Flesh | Open
The AI doesn't seem to be aware that it can safely block any enchanted creature. I kept attacking with them over and over and the AI wouldn't block if the attacking creature could kill the wall with damage.


#25 AI Improvement needed Resilient Wanderer | Open
It seems to at the same time:

- Be able to discard cards at any time during a turn, so it does.
- Not be able to pick cards which it discards well, so it doesn't.
- Not be the best at choosing which protection from color to chose (due to reacting to random casts/activations with it's ability and then choosing a random color due to neither being targeted nor attempting to be evasive - I assume)

Just a nice mix of typical mishandlings to note. Grotesque Hybrid does this too, triggers if the player is targeted and discards a seemingly random number of cards. I'll just go ahead and report this as a bug, honestly, I'm sick of submitting stupid reports on the bug tracker and irritating people, but wth, this is an actual bug. If it can't/won't be fixed or anything done about it right away, still worth a try.


#26 DECKBUILDER TIP - Pale Wayfarer | Open
The AI is very likely to use Pale Wayfarer ability just to untap Pale Wayfarer. So it's not likely to have it tapped when it needs the ability to save someone. I'm not sure if it's more likely to tap down it's mana to do this than play it's hand, but I wouldn't advocate giving it to the AI.

Also, it seems to make strange choices when picking the color. This, I think, might be related to why the AI is activation whichever ability grants protection from color. If it activares the ability just to untap the guy, it'll pick a random color it seems.


#27 Mistimed ability hazard - Cartel Aristocrat | Open
Cartel Aristocrat can, apparently, activate it's ability at enemy EoT. For no reason, too, judging by the random choice of color (blue vs. a non-blue board and a non-blue deck).

From the "gain protection" perspective, unprovoked enemy EoT is probably the worst possible moment to do it. From the sacrifice outlet perspective, possibly no so, but it's possible that that should be organized in a different way - a card which wants to be sacrificed looking for an outlet at opportune times, rather than the outlet firing off madly.


I'll add more if I run into them. If any of this looks properly reportable go right ahead, some of these are bound to be proper bugs.
---

My Shandalar deck pack folder is avaliable here:Dropbox
Leave feedback on particular decks here: Google doc
Ask for instructions, give feedback and complaints here: Thread
User avatar
lujo
 
Posts: 557
Joined: 20 Nov 2013, 13:17
Has thanked: 224 times
Been thanked: 70 times

Next batch (Still HS2)

Postby lujo » 14 Nov 2016, 18:21

The previous entries were for cards with the "protection" ability, this one will be a bit more random.

#28 AI Improvement needed Hearthcage Giant | Open
The thing with Hearthcage giant is that the AI doesn't seem to realize it only pumps giants rather than any creature. So once the AI plays Hearthcage Giant it will attack suicidally with it's other guys even if the settings are set to make it very conservative while attacking.

This, unfortunately, seems to make giving the card to the AI a bit of a hazard.


#29 AI Improvement Request Afflict and Zap | Open
AI is way too likely to throw them out randomly without killing anything. The whole system where the AI values the card drawing card by the value of the card about to be drawn is wrong on multiple levels and it shows in practice.


#30 Timing Stoneshock Giant | Open
The AI seems to wait untill EoT to activate the monstorosity ability of Stoneshock Giant.

The first problem is that it activates it at it's own EoT rather than opponents. The second is that it the ability should much rather be activated before attacking. The third one is that I've just seen the AI activate it when the Stoneshock Giant itself had summoning sickness, which is a bit of a waste as it can't attack then.


#31 Deckbuilder Heads Up - Cards with repeatable activated abilities pretty much only for the player | Open
Just don't use them in AI decks. #1265 Apparently just about any ability can be used in response to anything, which obviously means that putting more things which can prompt response in decks just invites mayhem, as the list so far shows.

This makes sense if cards were mostly ported with bulk and speed in mind rather than the ability of the AI to use them, which they undeniably were. This is simmilar to what I noticed about pump knights, but it obviously really does apply to basically anything which doesn't tap to do whatever it does. Much of that also seems to run on whishful thinking but is probably at least worth checking out.


#32 Bugged/Fatal Systemic Flaw - Dwarven Bloodboiler | Open
The following behavior likely extends to other things. The AI can attack into straight-up pointless suicide if it has Dwarven Bloodboiler on board.

This happens if the AI calculates that it can trade with the enemy by using the effect mid-combat, but it doesn't properly understand that it's supposed to have an untapped dwarf in order to do it. So it taps all it's dwarves for the attack and screws up pretty reliably.


#33 Possible AI tweak needed - Mark of Fury | Open
I've just seen the AI have two Akroan Conscriptor on board. One had summoning sickness, the other didn't. The AI cast Mark of Fury on the one which didn't.

I think I'll report this, I'm curious if there is an actual appliable tweak.


#34 Phase Tweak Needed Keeper of the Flame | Open
The AI is using it's ability at it's own EoT instead of enemy EoT. It will do this even if it means it will die due to not being able to block.

It's a VERY common thing with activated abilities, and it makes the AI seem crazier than it actually is (and it makes comparatively more bizzare behavior more difficult to spot).

I wonder if this is sloveable. If it's not it's another good reason to be vary of giving the AI anything with activateable abilities, but any kind and not just repeatable ones. I can't remembr what happened to reports of this kind of behavior it might have been fixed in the past and it's too late in the night for me :(
---

My Shandalar deck pack folder is avaliable here:Dropbox
Leave feedback on particular decks here: Google doc
Ask for instructions, give feedback and complaints here: Thread
User avatar
lujo
 
Posts: 557
Joined: 20 Nov 2013, 13:17
Has thanked: 224 times
Been thanked: 70 times

Re: AI Improvement suggestions

Postby lujo » 16 Dec 2016, 08:46

Right, Gemcutter's Guild from now on!

#35 Deckbuilder Heads Up - Don't give Cycling to the AI | Open
This was predictable, but still needs to be noted - the AI will madly cycle anything as soon as it can, or so it seems. For an example, if you give the AI a set of cycling lands like Tranqil Thicket, it's very likely they'll always be cycled, often during upkeep, and this will tie up AI's mana AND result in it getting mana screwed.

The behavior is complicated further because if the insane method of computing action value is still what the AI is using, then how much the AI wants to cycle something can and will be influenced by the value of the card it is about to draw rather than anything that should be influencing the decision. If this is the case then nothing with cycling should under any circumstance be given to the AI becuase it's impossible to reasonably predict what it will do - AI can be given cards it misuses horribly if it does so in a consistent way , but card draw which modifies it's value based on what's about to be drawn (with values including parameters like "based on it's rarity") is simply unusable.

So on principle, most cycling cards probably shouldn't be given to the AI at all. The possible exception might be if you specifically want the AI to always cycle certain cards which cycle for an effect, and if you can adjust the deck's curve to make it so that if the AI does this it doesn't affect anything. Like giving the AI Decree of Savagery knowing it will probably always cycle it at 6 mana and counting on that (although knowing the AI I can somewhat safely predict that it would use it to give an enemy creature counters if it has no creatures on board making that a rather bad idea).

How to fix this, and also basically anything else that involves drawing a card - probably a) have parameters for cycling not also be paramters for universal card draw so that the likelyhood if it being done can be adjusted and b) have a system of handling universal card draw which is related to and compatible with magic the gathering card design and assumed gameplay.

I'm also nothing it here, because I assume it won't be fixed, and there's likely no reason to even bother with trying to get it fixed. Knowing it's largely unusable could save any deckbuilder a bunch of time and energy, though.


Clarification and Update on #35 - I have tried out several more cards which cycle for abilities and the ones I tried didn't end up being cycled madly like the lands did. Oddly enough some seemed to be played better than pump generally does (which might mean that either pump was improved or whatever Korath did with cycles-for-effects cards is an improvement over the general pump). Which means cycling is not necessarily a huge NO but rather depends on weird parameters. I suspect it might be the perceived value difference between the card being cycled and the card being drawn in case of cards that cycle for no extra ability, meaning those might get cycled in horrible ways, while others might not.

#36 Deckbuilder Head's Up - Careful about modal spells | Open
This was predictable, but I had to try anyway, and yes, giving the AI modal spells is asking for trouble. This might depend on which modal spells you give it exactly, but in general there's a fundamental problem with them.

The AI has serious problems with the "modality" of spells you wouldn't think of as "modal", like pump spells which can be cast at any time, or burn which can serve as both life damage and creature damage. It tends to only play reasonably well with spells which are generally and historically considered too limited to ever put in a deck (Righteousness , for example.). If you give it something which has even more diverse applications, the chance of it doing something counter-productive or plain insane goes up.

Since it's allready somewhat difficult to predict what'll prompt the AI to do what, if you give it Seedling Charm in a deck full of deathtouch, you're as likely to see it play Seedling Charm at the end of it's turn to give it's 2/2 Trample for no sane reason as you are of seeing it do anything else with it.

The situation isn't necessarily as bad as the one with anything that involves card draw, regular pump spells, burn and such, but if you would give the AI a modal spell, you'll have to test the deck a bunch just to make sure it isn't getting routinely misplayed in it, and you also can't depend on it being used for what you want it to be used.

As for concrete suggestions for how to improve the AI - well, I've got nothing because it's not about any one particular thing, just the fact that many of these cards give more opportunity to the AI to suck at what it already sucks at. Requests to work on the AI of basic effects get "won't fix" for various reasons even on non-modal cards, so it's safe to assume many of these are player-only for the forseable future.


#37 One-Shot Regeneration Effects | Open
I'm not entirely sure but the AI seems to be reluctant to play one-shot regeneration effects. So much o that it looks like something about them is either bugged, disfunctional or missing. I've been noticing this for several versions now, IDK if they ever worked right.


Just a video illustration of #35 . If anyone thinks the consistent cycling errors are not very important, the difference in the final case was two attacks with a 5/7 guy, one on turn 4 and one on turn 5 (which is a dead opponent in starting out Shandalar) on top of effectively skipping a turn for no sensible reason, to say the least. If the card draw evaluation system is what's behind it - well then think of it as "Why the system is horrible video #1". There's bound to be more, I'm having trouble understanding what possessed people to even go with that system. Whatever it was, the benefits are inevitably minor compared to all of what it completely screws up.
---

My Shandalar deck pack folder is avaliable here:Dropbox
Leave feedback on particular decks here: Google doc
Ask for instructions, give feedback and complaints here: Thread
User avatar
lujo
 
Posts: 557
Joined: 20 Nov 2013, 13:17
Has thanked: 224 times
Been thanked: 70 times


Return to Shandalar

Who is online

Users browsing this forum: No registered users and 3 guests


Who is online

In total there are 3 users online :: 0 registered, 0 hidden and 3 guests (based on users active over the past 10 minutes)
Most users ever online was 279 on 11 Jul 2013, 22:03

Users browsing this forum: No registered users and 3 guests

Login Form