[gurps] Failing autopilots and RVO

[Zan Lynx]

Knapp wrote:
>> But humans can reach beyond their training when they need to, or sometimes they make a > mistake which later turns out to have really been the right thing to do.
>> That's less likely for a computer running a program.
> 
> Why?
> 
>> I am thinking that a computer which rewrites its own programs based on
>> experience or predictions of opponents future actions is no longer just
>> running a program; at that point it is at least an LAI (in my opinion).  And
>> that is a whole new thing than a Skill 17 piloting program.
>>
>> --
>> Zan Lynx
> 
> They don't rewrite anything and they are just running a program. They
> just find the best path (or a good path) to a solution and then
> remember it in similar situations. I feel that you are not getting
> what I am trying to tell you. It you watch the videos on the links I
> posted? Here is the same stuff from a different writer. Maybe that
> will help. Also remember that this is from 2006, 3 years ago. And from
> a talk about Go AIs.
> 
> http://senseis.xmp.net/?MonteCarloTreeSearch
> 
> "A playout is the basis of all Monte Carlo methods. It is a fast game
> played with random moves from a starting position to the end of the
> game, generating either a score or simply a result (win/loss). A
> playout can be light (completely random moves) or heavy (moves are
> biased based on heuristics, such as pattern libraries, shape, group
> status, move history, killer moves, etc.). (See also: playout
> analysis)
> 
> Win rate is a statistic kept by nodes in the game tree, tracking how
> many of the playouts from this position resulted in a win for the
> color to move. Research so far finds that this is a more effective
> statistic to optimize than average score. The use of winrate instead
> of score results in the characteristic Monte Carlo behavior of winning
> games by small margins. "
> 
> This stuff is all CURRENT, it is not some super AI. It can be done now
> and IS being done now.
> 

I am not sure you get it.  I KNOW ALL THIS STUFF and I still disagree 
with you.

If your Trade Federation is building 100,000 killer droids, is each one 
going to be trained individually, or do you think it's a heck of a lot 
more likely each one will have THE SAME initial database of "pattern 
libraries, shape, group status, move history, killer moves, etc."

That database is the program.

And even more likely, that the selection critera will tend to cause "the 
program" to converge on similar tactics over time, or at least wander 
between common "attractors", no matter how much training it gets.

Perhaps I did phrase my rating of an LAI inappropriately.  The software 
in your example is still limited to modifying numbers within its 
framework, it is not rewriting the framework itself.
-- 
Zan Lynx
zlynx at acm.org

"Knowledge is Power.  Power Corrupts.  Study Hard.  Be Evil."