Doomsayer and the A.I.

[Muppet1]

The card the A.I. struggles with the most is Doomsayer. It seems to pay no attention to this minion either on its side or its opponents and will happily play minions that are doomed.

Something like a don't play minions(that aren't totems), with no deathrattle or jade effect if an opposing doomsayer is staying in play this turn. Is I think a start.

[webadict]

Even more than that, sometimes AI will have enough to kill a doomsayer and NOT kill it, even going as far to deal damage to it, but not finish it off.

[doctorpangloss]

The AI doesn't look past the current turn and therefore cannot model the effects of Doomsayer.

[demilich1]

That is exactly the reason why the A.I. cannot "understand" Doomsayer

[webadict]

Well... Since I have to make a "Skip a turn," we could add an additional look at the start of next turn, which is probably helpful in determining priorities for other minions too.

[Muppet1]

Ah I see I sort of assumed some sort of Monte Carlo sim was used like in Magearena.

[doctorpangloss]

@Muppet1 Based on my reading of the literature that specifically uses metastone as an AI platform, it looks like GreedyOptimizedTurn is better performing that the MCTS already written in metastone. As lots of people have tried implementing MCTS AIs and none seem to be particularly newsworthy, there are a few things that could be true:

• MCTS does make it possible for the AI to understand this card, but it performs worse in aggregate against other kinds of opponent AIs.
• Hearthstone generally has so few "next turn" effects beyond understanding the contents of your deck, so accounting for these few cards makes the AI perform poorly with the sorts of decks that are randomly generated or that people actually play.

The first is almost definitely true. The second is probably just a characteristic of Hearthstone being a more accessible game than Magic. Looking at only your current hand is a really effective basic strategy in Hearthstone.