Can devs give overview of how AI is implemented?

[flupo42]

I have only discovered Forge a few weeks ago and having fun exploring the quest mode and what one can do in it.

Am very curious as to how you guys are implementing the game AI - I've been browsing threads in the developer section but not sure how relevant all the discussions from 2014/2015 are to the game today.

I see there are currently about 3.5k cards in the game that are marked as not playable by AI - can you guys give any info on where the challenges are with most of those?

Is it a matter of no-one having the time to script those cards for AI, or do you use some global generic algorithms that can't handle particular mechanics...?

(for example I noticed that a lot of the cards with that flag are those dealing with non-conventional costs/downsides to playing the card, like having to sacrifice permanent to use ability or maintain it in game)

How likely does it look at this point that AI will be able to handle more of those cards in the future?

How do you guys get AI to handle combat tricks and how does it plan (if at all)?

Are there any major improvements to it in the works?

[friarsol]

The AI isn't my focus on the project, but I can give you a little bit of info.

Cards the AI is marked for -

These cards are typically things that the AI makes terrible decisions with (paying costs it shouldn't pay, activating at times that prevent it from doing other more useful actions), with the worst case being cards that the AI does illegal things with. Honestly, I doubt all of the cards marked for the AI are bad enough to be marked. There are plenty of them that the AI uses just fine in Quest decks that are built around. If people are interested in doing an audit on all of these cards we might be able to "unlock" them. We try to keep all of our mechanics as generic as possible, so they can be used across the board. Some cards get some special treatment in some cases.

I'd say it's very likely that the AI will handle more of those cards in the future. When in the future? That I can't really answer, the more developers that are interested in Forge, the more likely someone will want to help the AI play better.

Umm... I don't remember how the AI does combat tricks, but I remember when we first added the capability a few years back it surprised a bunch of people who weren't used to it. The basics were something along the lines of "if this effect can be played at instant speed, and there's no reason to do it right this second, and it might be better to do it a little bit later. Don't run this as a possible action choice"

The AI doesn't really plan. One of the developers had come in playing around with things that could do that, but I haven't seen that be worked on in a while. I think (one of) the main hurdle(s) for that was determining values for game states. Hopefully he comes back and has sometime to work on it, it'd be nice to see some progression here.

[flupo42]

Is there any point in compiling reports of wrong decisions by AI using cards marked as unusable for AI - like to save time for any dev that might be looking to fix some of those?

[KrazyTheFox]

I'm not the developer of which friarsol spoke, as that particular project is more recent than mine. I started a different attempt at improving the AI that involves simulating future moves and predicting reactions in order to pick more likely moves. This is a huge project and it won't be done for a very long time. At the moment, it's at a bit of a standstill, though lately I've been thinking on how to progress the project. With a couple prototypes under my belt, I've identified some of the bigger hurdles to overcome (in particular, ensuring game states are synchronized, since every player, including the AI, simulates the game separately); I just haven't had the time to really sit down and plan things out. This would probably end up being more of a Forge version 3 than an improvement on the current version, as the entire game engine needs rewriting.