AI Improvements

[russianspy1234]

Here's a giant pile of fail by the AI, one of the biggest I've seen.  (and I've seen a lot :p)  With earlier precogs, I knew it had SoSas in hand.  After waiting out about 8 turns of SoSa, it didn't play any more of them.  It had them in hand for two full turns, and it could have stalled longer and decked me out, but it didn't.  No silences were played by me.  (I played a precog to make the screenshot better; I know better than to play them when my deck's that low against a stall :p)

AI doesn't play more than 6 SoSac in the same game. That's why SoSa deck aren't seen more often in Arena 

AI cast purify on my antimattered lobotomized purple nymph:

AI doesn't want your antimattered creature to die, less even if it's a nymph. That is used for good, antimattering momentumed dune scorps 
By the way, what's that beautiful timebow?

That would make sense, except my nymph wasnt damaged or poisoned., and there were plenty of targets that would have been good for the purify.
 I am playing a modified version of SGbow

Hover over cards for details, click for permalink

Deck import code : [Select]

4sa 4sa 4sa 4sa 4sa 4sa 4sa 4sa 4sa 4sa 4th 52n 592 5lm 5oj 5ro 5rp 5rp 5rp 5rp 5rp 621 6qq 6qq 6qq 6qq 6qq 6qq 6qq 6qq 6u3 6u6 71b 74b 74b 7am 7am 7dm 7do 7k2 7q5 7q5 7q8 7t9 80d 80h 8ps

Seperate game:
AI just cast purify on me multiple times.  It seemed to happen midcasts of ice bolt.


another game:
AI lobotomized my shrieker instead of my strong momentumed forest specter

[Belligerence]

I had AI Adrenaline an unubuffed Spark. It happened in an Arena game, and it was too quick to screenshot, but I did get the AI to do it again in Trainer.



I'll go out on a limb and say that this isn't exactly the most efficient use of Adrenaline.

Edit: I suppose with *only* Sparks out, this wouldn't be as bad (even though the Spark would die after the first attack anyway...), but the AI also had a Cockatrice and an Emerald Dragon, neither of which had Adrenaline.

[Chapuz]

I had AI Adrenaline an unubuffed Spark. It happened in an Arena game, and it was too quick to screenshot, but I did get the AI to do it again in Trainer.

Spoiler for Hidden:

I'll go out on a limb and say that this isn't exactly the most efficient use of Adrenaline.
Edit: I suppose with *only* Sparks out, this wouldn't be as bad (even though the Spark would die after the first attack anyway...), but the AI also had a Cockatrice and an Emerald Dragon, neither of which had Adrenaline.

I supose that the adrenaline code searches only for the highest attack creature with adrenaline effect, without looking the HP and poison on the creature. Another thing to improve.

[ponnuki]

The AI played Parallel Universe on its Spark (5/0), then went on to grow one of its other creatures to 6/6. It's doubtful already to PU a Spark when there are 'real' creatures to PU, but it's definitely a bad idea if one of those will actually be bigger than the Spark...

[fsk]

If the AI can kill me this turn by removing my shield, it should target my shield ahead of other permanents.

If the AI is close to victory, it should unburrow its shriekers.  The AI used to do this.  For example, I have 16 HP and the AI has two burrowed shriekers.  It should unburrow them and kill me, rather than dealing only 8 damage and giving me another turn.

The AI will only play Chimera if it leads to a kill this turn.  The AI used to play Chimera more often.

The AI should play its boneyard before killing one of my creatures.

The AI should sometimes save its enchant artifact for its weapon or shield, rather than always immediately playing it.

[FIQ]

I suggest that the AI, when considering buffing itself, should foresee exactly what the outcome will be at the end of the turn, i.e. by playing a turn in it's "head". This way, the AI would do less stupid things, I think.

[Chapuz]

I suggest that the AI, when considering buffing itself, should foresee exactly what the outcome will be at the end of the turn, i.e. by playing a turn in it's "head". This way, the AI would do less stupid things, I think.

Yep. And have a list of priority card and skill order usage.

EDIT: AI cremates ash eater with no other cards in his hand. dafuq?

EDIT 2: I was playing with an Instosis variation and played 2 supernovas in the same turn, generating a singularyty. That turn ended and the singularity gained Vampire skill. What did the AI do that turn? KILL THE SINGULARITY WITH A RAGE POTION lol xD

[Tirear]

EDIT 2: I was playing with an Instosis variation and played 2 supernovas in the same turn, generating a singularyty. That turn ended and the singularity gained Vampire skill. What did the AI do that turn? KILL THE SINGULARITY WITH A RAGE POTION lol xD

That's because singularity's high cost outweighs its negative attack. This may sound counter-intuitive, but it is actually a pretty good method of estimate the threat of an antimattered creature. Creatures with powerful abilities, such as purple nymphs, tend to have high costs, and the AI will recognize that 1 healing per turn is not worth letting them live. Creatures whose reason for their high cost is attack power will usually have attack >= cost, so the AI will leave such antimattered creatures alone. This saves the work of determining the exact value of each ability, and I think more kinds of CC could use something along these lines.

The real problem here is that singularity was not made to be played, and it was given a high cost solely to prevent the AI from playing copies received from nightmare. This had the side effect of making the AI assume it was a valuable card. This interaction goes to show why patches to the AI's behavior belong in the code of the AI. Teaching the AI not to play singularities would not only make it possible to lower the cost and solve this interaction, but also help AI decks that generate lots of entropy quanta.


In a similar vein, if you cast momentum on a deja vu* and then use its ability, the AI will never lobotomize the resulting creature, because it does not have an active skill. If the AI were taught to lobotomize momentum (the status), not only would this interaction be fixed, but you could get rid of momentum (the active ability), fixing the asymmetric interaction between momentum (the spell) and luciferin.
* Also works with lycanthropy, or using an unupped mutation to create an abomination, but buffing a deja vu before using its ability is a common strategy.

[Belligerence]

I don't normally comment on the AI playing multiple weapons on the same turn (or stealing yours then playing it's own), but this is just sad.

[asiantraceur]

I was playing a deck and it chained 2 Shard of Sacrifices...I don't use SoSac so I don't understand its mechanics that well but I believe what that results in is AI loses 63-80 HP without any extra bonus from the extra SoSac?

[Wizy]

I was playing a deck and it chained 2 Shard of Sacrifices...I don't use SoSac so I don't understand its mechanics that well but I believe what that results in is AI loses 63-80 HP without any extra bonus from the extra SoSac?

The turns add up. It is an ok play by the AI, that way it loses less non- quanta and avoids being countered by possible silences.
In certain situations it is more effective to wait until the previous shard wears off, but that's too much thinking for the AI.

[qwas296]

i dont know if this is already said but the ai just played bonewall and an instant dim. shield after that...