Re: [FT] Unpredictable AI

KH.Ranitzsch@t-online.de wrote:
> 
> David Griffin schrieb:
> > > It's at this point I hear two comments: 1) humans
> > > are "unpredictable";
> 
> > > In answer to 1: humans aren't THAT undpredictable.
> > > In fact, a lot of fields (marketing, social science,
> > > polling -- *ahem* --) are based on that fact.
> 
> I consider the frequently stated argument that combat AIs would be too
> predictable to be a red herring. Even today, it is a simple matter to
> write a computer program with a random number generator to make the
> program as unpredictable as you like. It is well known in game theory
> that for some games (e.g.paper, scissors, stones) a purely random
> strategy is optimal. Any decent combat AI would be able to analyze the
> chances of a possible tactics. Weigh the options according to this and
> pick one at random based on the odds calculation.

Humans are certainly more predictable than we'd perhaps like to think we
are. OTOH, machines are less so.

As some may know, my 'day job' - the one that actually pays some money,
unlike my Brigade activities :-) - is writing games software. I've
written the AI code for both fighters and cap ships in two PC games so
far. OK, I'll accept that this isn't 'real' AI but it's probably the
closest anyone on this list is going to get to writing starfighter AI
for at least a couple of hundred years !

One thing my AIs DON'T do is always pick the best option in a given
situation. They evaluate several options, weigh them up and then pick
one at random (fuzzy logic). These options will be weighted so that the
best option is more likely to be picked, but the potential is there to
not do the obvious. Any AI of the future will probably do the same, but
more so.

The danger in AIs is what they do in an unknown or unpredicted
situation. They could cope, or alternatively they might go haywire (I
know, I've been there - 'why the heck is that ship flying in circles
?'). Asimov's 'I, Robot' has some great examples of how rigidly applied
rules of logic don't always work.

Tony

-----------------------------
Tony Francis
Senior Software Engineer