About seven years ago, I became fascinated with an AI project called ''Cyc'' (www.cyc.com) and I began to learn about meta-level knowledge bases, frame based AI, scripts, etc. It got me thinking about the ultimate future of games, and I arrived at this: Just as a storyteller narrates a story and improvises when appropriate, imagine an AI ''gamemaster'' which dynamically plots the role of your game persona. In other words, the AI sets the stage for you, and lets you engage in play. As gameplay unfolds, the AI draws upon its vast knowledge of plot techniques and the gameplayer''s tastes to make a unique and custom tailored experience for the player. Each player would have a unique experience, and each time the player begins anew, gameplay would unfold differently. Essentially, what would be required in addition to advances in AI, is an AI which has knowledge about plots, schemes, roles, writing, etc. Looking from this hypothetical future backwards, what would be the first step to initiate gameplay like this? Taking into account that we don''t have this advanced AI technology, what would be possible now? It seems possible that we could develop a ''lite'' knowledge base of mini-plots with fill in the blank slots. I know that someone else mentioned this concept on another thread. The key is to program the AI with enough panache and knowledge of context so that it can make all these plots gel effectively. I also propose that the AI should be able to build a model of the player and their tastes and desires. The AI should keep notes on the average duration that the player plays, the situations in which the player is most responsive and active, and which activities the player appears to enjoy engaging in. Any thoughts on the ''future'' of AI gamemastering?

dude, making a game that taylors itself to the player. Would it be event driven? Like, you kill this guy and the story branches from there? This has been done before, but not very well and with not enough divergance. But to make the game change based on gameplay style, wow! I can''t even imagine it. Argh, my brain hurts. Okay, how about different values that represent a persons virtuosity. This could be done by measuring how often players kill instead of sneak or threaten instead of ask nicely. I don''t know...man, this sounds tough.

Bowowow yipeeyo yipeeyay

quote: Original post by capn_midnight

This could be done by measuring how often players kill instead of sneak or threaten instead of ask nicely.

This is along the lines of what I am saying. So if the player spend a great deal of time skulking around dark alleyways and hiding behind doorways, then the AI draws upon its knowledge of plots, settings, etc. to generate situations incorporating aspects like this.

However, the AI would also be experimenting with the player, giving him new scenarios to feel out how the player responds, and thus building an even more accurate model of the player.

Yeah....how to make an AI that works as a *good* GM. Hmm. I suppose you could force-feed a genetic algorithm program entire PnPRPG campaigns/missions. But then again, it''d probly just kill itself when it got to the AD&D 3rd Ed. Rules....
On second thought, thats a bad idea.....
OOOH! Stick some electrodes REALLY far into your ears and think about good storylines....mebbe the computer will pick something up....*grin*

This RtS-caffeine-babble has been brought to you by:
-Run_The_Shadows
-Run_The_Shadows@excite.com

OK, lets look at this in more detail. Envision several layers to a game program:

First layer:
The graphics engine, responsible for rendering and physics simulation.

Second layer:
The simulation engine, divided into several modules, each responsible for simulating either the economy, characters, weather, etc. This layer calls upon the graphics engine.

Third layer:
The AI GM, responsible for learning about the player, observing the whole process, and making sure gameplay unfolds in an interesting way. The AI GM may allow the first two layers to operate uninterrupted until it believes it is appropriate to interject something.

Here''s an example:
Visualize the Sonoran Desert, complete with cacti, mining towns, gunslingers, outlaws, saloons, etc. Your character enters a saloon and joins a poker game with some rough looking characters.
Everything is running under the first and second layers right now. The second layer simulator is handling the cards and providing a perfect simulation of a deck of cards, the NPC betting, etc. The AI is observing, but letting things run their course.

A few hands are played and the cards are reshuffled. The AI is evaluating the situation for possible ''interesting'' plot twists. At this point, the AI overrides the simulation and calls upon a poker game scenario. The AI notifies the simulator to stack the deck. Your character is dealt four aces, and the mean looking guy across from you is dealt a full house. Betting ensues. The AI has decided that if you don''t fold, the mean looking guy will snap, threaten you, and draw his gun, possibly creating a very dangerous situation.

Now, the above scenario cannot and should not just be randomly applied. Rather, the AI has noted its application, and knows that it should not apply this type of scenario again for a long time. On the other hand, if the player plays poker again tomorrow and the simulator creates a similar situation on its own, then that is just the luck of the draw.

I love the idea of having that kind of plans. But the question is do you have the time and the funding to achieve that project? This idea would have been great for Deus Ex!

DarkFerrett

btw, check out my post about my halflife modificatioN!

www.shogunmod.com
Its your time to choose.

I think the first step for a game like that would be to create a game scenario that the AI could manipulate. How many variables would have to be manipulated to create the situation with the cards? And how would you procede in a situation like that? Draw your gun and have a 50% chance of survival? Doesn''t sound very fun.

Think of how you would make a game, where if a human were to be the GM, all he''d have to do is click a few buttons to manipulate the game play. Then from there, you can write an AI to replace the human GM.

E:cb woof!

I think a GA would blend into this task sutibly(sp? )
get it do determine the events its watching for, and the scripts that it will execute if it finds that event, (within paramaters).
It could also change the script- but that will be hard(er)

Or maybe Minimax, you''d have to get the values right (GA to the rescue!), but if you get it to see which moves the AI looks for (in context. Eg. Tough guy comes in, other guy cheating), and possible reactions (Leaving bar, picking fight).

If you have a small databace of possible actions/reactions (per context, but a lot total). (possibly another set of atributes to ''kick in'' different contexts, possibly by objects- eg. There is a piano, the algarithm would add the relevent contextual actions to the pool, and remove them once something happens (you move too far away, it gets smashed, ect). You could get the Algarithm to get a good plot going. Though it will take some time. (it is a GA- it takes a Looong time to do things like this).

Nice Coder. Look at the Dates. You''re really bringing back alot of OLD threads that died a long time ago.

I''ve heard next generation of The Sim has something a little similiar to this awhile back but with user selectable content instead of full AI GM.

Players would upload a saved file of their little character''s behavior onto a Maxis website where there is some kind of AI script that would automatically profile all the data from players around the world and tries to match them in different situtations. Players could then define a scenrio of events on the website that they would like to see occur and the webserver would offer them in return another player''s profile for download and plays out the whole scenrio event in game.


>>I also propose that the AI should be able to build a model of the player and their tastes and desires.

Agrees with most of it but which user feedback but "tastes" is not a very efficent way to build a model on, or at the very least, hard to measure a user''s feedback; you can''t have all the npc in a town goes up to the player and asks "do you like my <style> cloth?"