On 03/25/2013 03:27 AM, Thomas Thrien wrote:
Say you have an AI trained to an impressive skill level and you want it to
self-learn to become better. Like any other character in that situation, it
would be learning without a teacher, without books, on the job. The GM should
ask the players just what that job is …
Not one AI, dozens, hundreds. Each time a device returned from the field it
exchanges its new experiences with all the others. The knowledgebase of Version
1 is used to create that for Version II, so that your Battlesuit Mk. VII can
look back on the experience of 200 years of infantery combat even on the day it
leaves the factory.
Confessed, that is the ideal world from the marketing leaflets. Part of the
game would be to find out where the sales blabla deviates from product reality.
The real world will certainly deviate from simulator training, that's
for sure. If the manufacture continues to incorporate real world
experience and also includes random elements in its simulators, the
training and skill increases will continue to be good and relevant.
One danger of having neural nets training others in a repeating cycle is
that they will learn and optimize the tactics that work in their little
world against their known opponents. A new opponent that uses a
completely off the wall tactic may confuse them. In game terms, perhaps
the AI or skill program has a familiarity bonus vs. other skill programs
or a big negative vs. low skilled opponents. (He's so clueless at space
fighter combat that I have no idea what he'll do next!)
Unless the algorithms jump off into random untried territory every so
often they can end up in local maximums. There may be a *much* better
approach but the algorithm won't try it because it has an existing
solution that mostly works.
These problems happen to people too.
_______________________________________________
GurpsNet-L mailing list <[email protected]>
http://mail.sjgames.com/mailman/listinfo/gurpsnet-l
