Re: An AI question and a UI question

On Mon, 1 Dec 2003, Jeff Lait wrote:
>
> The Sheep wrote...
> >
> > No, no. The single event is the shoot. Not whole encounter. The shoot
> > might hit or miss. When monster shoots and misses -- it's bad for him.
> > If he shoots and hits -- it's good. Then, the player doesn't stand still.
> > When the player shoots and hits the mosnter -- it's bad. When he shoots
> > and misses -- it's bad.
>
> 1) This seems highly asymmetrical.  If the monster shooting and
> missing is bad, surely the player shooting and missing is good?

  If you were walking down the street, or dungeon, or whatever,
and I shot at you, and I missed, would that be "good" or "bad"? :)


> After all, if the player can't hit me due to my superior AC, I
> *should* stay at a distance and keep hitting him.

  I think The Sheep said this, in the next paragraph (comparing
h2h and ranged attack values).  If the player is known to be a
terrible shot, then sure it makes sense to stay back.  But I
think the monster is supposed to be in the dark about the player's
average accuracy, and thinking more along the lines of "Holy ***,
that guy's got a bow!  Help!"


> 3) In any case, this is *easily* determined in close form.  There is
> no need to train a NN with the inputs of the success of hitting with a
> missile attack.  We can exactly calculate the MOBs chance to hit with
> the missile attack.  We can exactly calculated the expected
> damage/second of the MOB against the human.  If your goal is to muddy
> these values by having them trained by the NN, I'd suggest it's a lot
> more efficient to start with the real values & muddy those.  Then you
> have fine control over the smarts of the mob.  You could even do this
> artificially.  Every time a MOB strikes at the avatar, increment a
> learn counter.   The size of the learn counter can be used to feed
> into the AI by providing a more accurate estimate of the to hit
> chance.  This obviates the need for any complicated NN design with
> questionable convergence.  One can tweak the exact rate of
> convergence.

  I tend to agree, but then I don't know anything.  Besides, NNs
are so much *fun*!  (No, I have no experience with neural nets
either...)  ;-)

-Arthur,
luckily, he missed the pitchfork.