[MUD-Dev] Procedural Content Generation [was Design: Episodic MMOG's (not expansionistic)]

Sasha Hart sasha.hart at gmail.com
Fri Feb 10 12:33:33 New Zealand Daylight Time 2006

David Wright wrote,
> I was curious if there has been much thought or discussion regarding
> Episodic MMOG's, wherein the content is transient - ephemeral, but
> characters and various npc's are persistent (or re-born), instead of
> the traditional where most/all content tends to stay around forever
> and players/npc's spiral ever upwards.
> Obviously this leads into the idea of disposable content, however
> I'm thinking this may be a Good Thing as opposed to content (and
> bugs/design flaws) which Never Dies.

Sean Howard wrote,
> The fact that you'd need to constantly create significant amounts of
> content is perhaps the greatest knock against the idea, though I do
> believe that procedural content (ie random dungeons) that can build itself
> would significantly reduce that problem - were it done well, which hasn't
> happened yet (though I do believe I've made a breakthrough in my personal
> considerations, but explaining it would be off topic).

Sorry to muddle this thread, but this made me think: I've seen systems
where people basically choose the sort of dungeon they want generated
(each type is associated with a certain batch of parameters, tileset,
whatever) - has anyone played with the game collecting and using data
about the sorts of procedurally-generated dungeons specific players
choose over others, or the like, to tune the generation parameters for
dungeons offered to those players?

It strikes me as a matter of (A) exposing a meaningful and
not-super-high-dimensional parametric interface to the generation
routines, (B) using some kind of reasonably cheap optimization
algorithm to skew the distribution of choices partly, while still
offering enough variety that the player won't get too bored and will
still have the ability to switch. Alternately, such systems, with a
slow change rate, geared toward optimizing parameter combinations for
the whole game or particular segments, etc.

This is purely a blue sky idea. I reckon the more fundamental
limitation IS the basic quality of the output of the generation
routines and the usefulness of the parameters. In other words: whether
the generation engine is significantly better than hitting a pinyata
full of props and mobs.  But, supposing that you want to turn over
content generation entirely to computers basically for cost reasons,
one possibly important thing that they don't provide is revision in
the face of feedback. I could see some kind of ability to revise
what's being prevented in this kind of simple way doing a little tiny
bit to make this option more acceptable. Knowing this is the real
world, it would still be artificial and likely inferior, of course -
maybe like sucralose vs. saccharine - but since content is evidently
such a problem for all those games which can't simply co-opt the
Simutronics system...

Alternate version: I suspect if you took a reasonably scalable
supervised-learning approach (support vector machines, etc.) to
approximate, say, the mapping from parameter combinations to observed
player preferences, (as opposed to just hill-climbing around with
simulated annealing, genetic algorithms, etc. to find a distribution
of parameter-combinations which worked better than other explored
distributions), that you would get faster learning, more stability,
and more human-inspectable information about what the game was
learning about the design-space your parameters exposed. That could
also make manual nudges more feasible than the hill-climb approach,
because data collection can be decoupled a little more cleanly from
the actual policy followed in choosing what to offer to players.

But that's just a hunch, since I haven't done anything on this except
to play with hand-crafted procedural wilderness generators. You would
need to make some assumptions about preferences in order to interpret
the data, with a big impact on your results. Ratings could be fairly
obtrusive, but anything else is a little less obvious. I've talked
about choice, but to some extent players won't be able to tell whether
they'll like it before they choose it... Stay time? Well... I would
consider ways to penalize dungeons for lack of completion, and to
expose dungeon properties in the quest description etc. (e.g. player
hates urban escort missions seems like a sensible thing to expect of
this). You would also face the possible headache that motivated
players will naturally choose or rate highly anything which enhances
their ability to treadmill faster, so if you wanted to control that
you might want to make your preference observations relative to rate
of return. Otherwise, even if your generator tries to ensure rough
equality of rate of return via prize density etc., your game will
happily work together with the players to mercilessly exploit its own

I don't personally doubt for a minute that getting smart people to
write zones for you requires less dev time and will have better
results (and can probably be done without paying them, since that can
be fun too). Maybe it's just another thing for dilettante hobbyists to
play with. But given the premise that you want to do PCG and to get
good results, it seems at first glance like an unusually well-defined
problem to which pretty standard techniques can be applied, and which
might free you from having to hand-craft the PCG parameters, make
adequate adjustments when the rules change, etc.

MUD-Dev mailing list
MUD-Dev at kanga.nu

More information about the MUD-Dev mailing list