D&D General Has anyone seen this Wired article about using D&D to teach AIs?

[Dannyalcatraz]

Oh great, now that AI's are getting involved how am I going to continue my high D&D win percentage? Technology ruins everything.

Use the Cyrano Gambit. Have an AI on your side, advising you via texts or your earbuds.

 

[Dausuul]

Interesting approach there, but it's not how I would have gone at it. Creating a coherent narrative is one of the highest-order functions of a Dungeon Master. It's like the Wright Brothers trying to design a space shuttle.

If I were creating an AI DM, I would start with the most mechanical function of a DM: Applying, interpreting, and extending the D&D rules to cover all the weird stuff players want to try at the table. That's a much simpler task than crafting a narrative, and much more amenable to the way modern AIs are trained. You'd have to put together a big dataset of "stuff players tried and how the DM ruled it should work," but that is at least in theory possible.

From there, one could proceed to creating an AI DM that can run a combat encounter in isolation (including appropriate tactics for the monsters, which does not necessarily mean "optimal").

Then you can start figuring out how to feed it one of WotC's big adventure books and have it run that adventure. Making homebrew adventures is the final step, not the starting point.

 

[Maxperson]

I did see this a few days ago. I have a feeling that an AI DM would turn out something like a Madlibs. Not that that is a bad thing it might make for some memorable game sessions. Now Im curious how my next game would turn out if I secretly kept a Madlibs page behind the DM screen and used it for improv during roleplaying scenarios.

The Bugbear is coming at you with a hairy screwdriver!
.........Monster.................................................Adj.... Noun

 

[UngainlyTitan]

Interesting approach there, but it's not how I would have gone at it. Creating a coherent narrative is one of the highest-order functions of a Dungeon Master. It's like the Wright Brothers trying to design a space shuttle.

If I were creating an AI DM, I would start with the most mechanical function of a DM: Applying, interpreting, and extending the D&D rules to cover all the weird stuff players want to try at the table. That's a much simpler task than crafting a narrative, and much more amenable to the way modern AIs are trained. You'd have to put together a big dataset of "stuff players tried and how the DM ruled it should work," but that is at least in theory possible.

From there, one could proceed to creating an AI DM that can run a combat encounter in isolation (including appropriate tactics for the monsters, which does not necessarily mean "optimal").

Then you can start figuring out how to feed it one of WotC's big adventure books and have it run that adventure. Making homebrew adventures is the final step, not the starting point.

Hmmm! not sure about this approach. Do you base the AI's adjudication on real world physics or "it would work in the movies". I think the latter would work better but I could easily see the main problem being edge cases where the AI's conclusions over stress your Suspenders of Disbelief.

 

[Dausuul]

Hmmm! not sure about this approach. Do you base the AI's adjudication on real world physics or "it would work in the movies". I think the latter would work better but I could easily see the main problem being edge cases where the AI's conclusions over stress your Suspenders of Disbelief.

That isn't how AIs are programmed. They aren't programmed in the way we normally think of it. Rather, they are trained; you have the AI try stuff, and reinforce behavior that leads to the desired outcome while discouraging behavior that doesn't. (Obviously, this is a drastic simplification of the process, but it's the basic idea.) Over time, the AI learns and gets better at its task.

In this case, the desired outcome is "human gamers feel the AI made a good ruling." So you need to find a way to test the AI's rulings for how human gamers regard them. One way to do that is to collect a giant dataset of oddball rules questions, get a bunch of experienced human DMs to issue their rulings, and then reinforce AI responses that resemble those rulings*. Another would be to incorporate it into D&D Beyond, where you can ask the AI for a ruling, then rate the answer it gives you.

*Before someone suggests it, Sage Advice would make a lousy dataset. Crawford has this weird obsession with sticking to the Rules As Written rather than suggesting practical rulings or giving guidance on the design intent. Rigid adherence to RAW is not generally a good trait in a DM.

 

[slobster]

Yup, AI can finally do everything we can do, but better. I went over to one of the neural nets and had it generate an adventure plot for me based on the first two sentences from Amazon's blurb for Descent into Avernus. The computer GM's pitch was as follows (bold parts taken from the module's own description to seed the algorithm):

"Welcome to Baldur’s Gate, a city of ambition and corruption. You’ve just started your adventuring career, but already find yourself embroiled in a plot that sprawls from the shadows of Baldur’s Gate to the front lines of the planes-spanning Blood War! In a world that, at times, is divided into four parts:

1) The North and the South
2) The north, the North, the South and the South.
3) The North and the South.
4) The North.

What’s going on here? How is Baldur’s Gate up to the North? You have to play through this to get it all the way to your destination. . .In the Middle of the Sun.

The North is a world of intrigue, intrigue, intrigue, intrigue and intrigue. It’s a big, complicated world, and all the secrets in Baldur’s Gate are hidden under a very mysterious layer of mystery."

naughty word guys, I think Skynet nailed it.

 

[UngainlyTitan]

That isn't how AIs are programmed. They aren't programmed in the way we normally think of it. Rather, they are trained; you have the AI try stuff, and reinforce behavior that leads to the desired outcome while discouraging behavior that doesn't. (Obviously, this is a drastic simplification of the process, but it's the basic idea.) Over time, the AI learns and gets better at its task.

In this case, the desired outcome is "human gamers feel the AI made a good ruling." So you need to find a way to test the AI's rulings for how human gamers regard them. One way to do that is to collect a giant dataset of oddball rules questions, get a bunch of experienced human DMs to issue their rulings, and then reinforce AI responses that resemble those rulings*. Another would be to incorporate it into D&D Beyond, where you can ask the AI for a ruling, then rate the answer it gives you.

*Before someone suggests it, Sage Advice would make a lousy dataset. Crawford has this weird obsession with sticking to the Rules As Written rather than suggesting practical rulings or giving guidance on the design intent. Rigid adherence to RAW is not generally a good trait in a DM.

I am aware of how modern neural net AIs are trained and am unsure how the goal "human gamers feel the AI made a good ruling" is less ambitious than the original goal. Games, with real people break up over issues of players and DM not being on the same page as to rulings and that is with people that have evolve to do social. As for the dataset, streamed games that last a campaign arc is probably a good place to start. Whether there is a sufficiency of data or how one could convert that information in to a training data set is a whole other ball game.

 

[UngainlyTitan]

I should add to the post above, that using streams as training data will probably skew the AI to play styles that are entertaining to watch, as distinct to entertaining to play.

 

[Mallus]

This can only end in AIs who try to set everything on fire!

 

[Dausuul]

I am aware of how modern neural net AIs are trained.

Ah, sorry - so many folks talk about AI as if it's just a fancy form of traditional programming, with the programmer scripting out a decision path. I should not have assumed that was what you meant.

...and am unsure how the goal "human gamers feel the AI made a good ruling" is less ambitious than the original goal. Games, with real people break up over issues of players and DM not being on the same page as to rulings and that is with people that have evolve to do social.

It's certainly not unambitious. But a table ruling is much simpler than a whole story arc. Generating a couple of sentences is far less taxing than generating multiple paragraphs, and it seems more feasible to build a training dataset. One approach would be to start with questions that have clear answers in the rules: "I rolled a 17 on my attack against AC 15. Do I hit?" "How many damage dice do I roll for fireball?" Then get it to where it can process a new spell or feat or monster that it hasn't seen before, and answer similar questions about that. Then start working on more ambiguous stuff.

I'm thinking of what the DeepMind team did with AlphaStar: They tried just letting the AI teach itself from scratch, the way they did with chess and Go, and it couldn't get beyond worker rushes. So they broke down the game into more manageable chunks and trained the AI step by step, starting with how to gather resources. Table rulings seem like the analogous place to start for a DM.

As for the dataset, streamed games that last a campaign arc is probably a good place to start. Whether there is a sufficiency of data or how one could convert that information in to a training data set is a whole other ball game.

I hadn't thought about streaming as a source, but that does seem promising--if, as you say, there's enough data.