Re: Bag of Tricks: Choosing actions via the matching law

Jesper Mathiassen wrote:

> On Mon, 03 Nov 2003 11:07:24 +0000, Eternal Vigilance wrote:
>
> >
> >
> > baylor wrote:
> >
> >> Problem Area: action selection
> >>      AI Type: decision making, learning (secondary), personality
> >> (secondary)
> >> Detail Level: mid-level
> >>    Technique: matching law
> >>  Assumptions: Options are relatively equal
> >> Example Uses: sports: choosing a shot type
> >>               FPS: choosing a weapon
> >>               RPG: choosing a spell
> >>               RTS: choosing a build unit type
> >>
> >>
> >
> > How many games these days have no additional modifiers that effect the
> > success/failure  (ie-whether is underwater,  or if char has his 'Hat of
> > Kiick-Ass Flamestrike on, or  orcs have a genetic diomorpohism resulting
> > in some having thicker heads - less success of StoneStorm...))
> >
> > If they are only a few factors, this method could be applied using
> > seperate
> > statistical data sets  for each classified situation flavor.
> > (Orc + Hot environment + no attack mods etc...)
> > Of course if the classification is too specific, then gathering enough
> > statistical data
> > to adjust the preference would become a problem.
> >
> > But then just doing it anyway and ignoring strong modifiers would be
> > an improvement over the way many of the current games do things
> > ( a standard attack preference with fixed probabilities for a particular
> > target type)
>
> Interesting thread this :) so for once, I'll contribute...
>
> Of course, these statistics would only work in specific scenarios,
> however, if the scenarios weren't somewhat different the wizard should do
> no other thing but to examine which situation it was given and employ the
> most powerfull attack. No adaption would be required.

Its funny that the mechanism to classify the situation is a magnitude (or
more)
complicated than the basic (simple) statistic weighting.


>
>
> However, if the Wizard had no means of determining anything about the orc,
> forehead mutations for instance, and say 20% of orcs would have this
> feature, this would in fact be what the statitics would show. So, the
> wizard would use Stonestorm in some cases, knowing that this would fail in
> some cases, but in others it is very powerfull. Choosing whether to use
> this attack could only be made either at random, or by looking at
> statistics. Statiscis would likely be the best bet.
>
> I beleive Reinforced learning often deals, or can deal, with these
> problems.
>
> - Styx

A more complicated mechanism might be needed to  figure out what to try next
after that first  Stonestorm fails (depending on if the game mechanics
indicate failure
of the attack as seperate from spellcast failure).

I was just thinking that the system has to be able to 'learn' by gathering
data
across events like player/char death (which in RL usually terminates the
learner...),
and probably would be useful to get larger statistic counts by observing other

similar events/results.  (NPCs might also get a leg-up by watching what
Players do
in a similar situation... )

Does the system become unworkable if there are too many options to test
(not just 3 spells but many more ).   How many test cycles then have to be
gone
thru before the best options start emerging in the data....

Hopefully the race of Orcs isnt made extinct  before the learning system can
adjust itself.