Could We Teach a Computer to GM?

After having toyed with the idea for a while, I think yes, but of course, its not quite that easy.

Of course, some of you will be thinking that your special beautiful flower of a game is not something a machine could come up with. Perhaps so and some of you will be right, but most of you won’t and that’s not the question anyhow. The question wasn’t whether computers could GM as well as the best humans. This is unmeasurable anyway. This isn’t chess. The question was whether a computer could be taught to GM.

First of all, if you run games as an “auteur”, basically showing off to your players, while they follow the tracks you lay down for them, a computer could do that easily. Even your story isn’t as creative as you think and computer could probably come up with it as well, given enough information to base it on, but this isn’t the interesting part.

Back in the 60s, someone (actually Joseph Weizenbaum) came up with ELIZA, an automated therapist that’s still part of most Emacs (a UNIX text editor) releases. It recognizes patterns from sentences and carries on a conversation based on that by asking questions. You could say its just a chatterbot (and it is), but apparently it still fascinates people after almost fifty years, so there must something more to it. This approach has actually been used in Dungeon about four decades ago.

A few years back (2011 to be precise), Watson won a special episode of Jeopardy!. Watson is a system built to answer natural language questions. It has huge databases at its disposal (the version on Jeopardy! had – among other things – the complete text of Wikipedia for it to search through) and it can search relevant words as well as parse together meanings from context while doing so.

Well, obviously the former is easily available to any of us and the latter is currently well out of reach for any but the richest of us. But how is this relevant?

My approach to GMing these days is ‘trust the players’. I want to leave as much as possible to the players. My role is build a context and keep the flow going. Players will entertain themselves and each other, if given a chance. I’ll happily ask my players to find complications for themselves and they are happy to do it. Its a matter of honor to find interesting outcomes for failures or what they miss on their partial success. I don’t need to be very creative, if the players are handling that part for themselves. Could a computer do this? Even now, probably quite well.

Computer can even go deeper. There’s a model known as BDI (belief-desire-intention) often used in AI research. Belief is what the AI knows about the world, desire is the goal its currently working on (which can change dynamically) and intention is the steps it plans to take to achieve goal. However, there’s another reason to use this model: when trying to model others. The GMing AI could build one of these for all important NPCs, but also for each PC.

This is a commonly used tool in many systems, which require adversarial reasoning, such as automated poker players on the net, but its also used in combat simulations by various armies around the world. This approach can be used to make excellent bad guys. If they have their own BDI, they can do stuff behind the scenes and leave a trail. The players could fight him on different axis besides the usual physical confrontation and the computer could handle all this quite easily.

All in all, I’d say a computer could (with current technology, but probably not technology available to anyone) GM very interesting and deep games, with multifaceted plots and a realistic world.

We do have problems though. Parsing speech in a normal social situation is pretty complicated. If we have a pretty normal four players who at least at times talk over each other, its very hard to follow, especially for computers, although its a problem that can probably be overcome (and has been in the past by using MUDs).

Its also hard for a computer to read social situation. When has some discussion gone far enough? When should you interrupt an in-game argument? You could probably find a way to count how many times a certain theme is repeated and if its too many, you should make the players move on.

Then there’s the rules of the world. If there are disagreements among the players on what is possible in the world, the computer might be unable to act as the final arbiter in such a situation, especially if we are discussing magic or technology currently out of reach of humans. This can probably be best overcome by simply limiting the options, or just relying on players to use some system to determine all this beforehand.

Leave a Reply