[gurps] Failing autopilots and RVO

[Knapp]

> Ok, I see you are wedded to your Monte Carlo idea.  I won't even argue
> against it, it is an appropriate algorithm for many things.  But it is still
> predictable!

I am not wedded to it. Just picked it because it is easy for most
people to understand. Lets look at it from the perspective of a path
finding algorithm. You start at A and want to get to B. There are
obstacles between A and B. You would start by going in 100 different
directions and 100 different distances (together not separate) then
repeat until you get to the target at B with say 10 of them. Then you
pick the shortest path of the ten.  If you did it for 1,000,000 your
answer would be better but take longer or use more computer power.
Using something to make it a heavy might also lead to it being more
predictable but more efficient but could also blind it to some
possibles.

You can not predict the path the computer will take nor can you say
that it took the best path. All you can say is that it will find a
good path that will be different each time. Here I am only looking at
a light Monti Carlo.

> It picks a random possible solution and then tests it and tries to move
> toward a local optimum solution from that starting point.  Those equations
> that define optimum solutions are themselves fixed!

They could be.

> So a program using Monte Carlo simulation will quickly pick some answer, but
> certain solutions will obviously have equations that make them *better* than
> other solutions, and the decision making will favor those.  As it should be!

Yes, as in my example the path goes from A to B and it picks the
shortest one but that does not mean that it picks the same one each
time. Even if it did 1,000,000,000,000 tests, it still could find a
bunch of paths that are the same length. Give one obstacle going right
will give a path just a good as going left around an obstacle exactly
between the two points A and B.

Add a bunch of obstacles then the possible perfect paths goes up with
the square of the number of objects, given perfect geometry of the
position of the obstacles on the path. The path picked would be random
and not predictable but there would be statistical probability of the
walker being at any give point at any given time. Also given that the
path picked is random, it most likely will never be the perfect path
but randomly off.

> An opponent that has the Monte Carlo decision code in hand or has observed
> it over time will know what the program is likely to do. Every copy of the
> program will have the same probabilities.

Just as a good soccer player knows what the other team players are
likely to do. This is just war as normal. Also we were at first
talking about freighters not fighters, right? We are talking about
handling emergencies as good or better than a human could.

> Ok, there could be several "personalities" defined for the program, but each
> one will have known quirks.  Or the criteria could be adjusted at random,
> but I hardly think that's going to result in an effective program.

Depends on how it is done. It could be effective is a sort of
aggressive VS defensive sort of way. People have the same problems.

> I am a software developer, a "hacker" who loves computer stuff.  I *have*
> poked around in the code of several Quake 3 bots, poker and chess opponents,
> so I do know pretty much what I am talking about.

Take a look at the modern go programs. They are the real current
cutting edge. I think poker is too.

> There are definite limits to randomness.  For example, one Q3 bot used
> randomness to select choices from decision trees, a kind of state machine
> thing.  It had some override code in there to keep it from jittering in
> place.  Deciding to charge forward or retreat at random is all fine but if
> you change your mind too often the effect is of standing still in one place.

True, but totally primitive!

>LAI

What is the L?

> Zan Lynx

I am hear for fun too!

-- 
Douglas E Knapp

Why do we live?