AI warrior.

>
> The 5 instruction limit make that hill makes it a good place for
> evolvers, including the human based variants.
>
> 'Neural Networks' - wrong hammer for this nail.
>
Thank you! I will investigate more deeply.....
However there was a misunderstanding:
with "warrior using AI techniques" I meant
"warrior whose code implements some AI
inside, like fuzzy logic".
 >> Stay informed about: AI warrior. 

On 27 Jul 2006 07:08:54 -0700, <cristian210273.TakeThisOut@yahoo.it> wrote:
>
> Thank you! I will investigate more deeply.....
> However there was a misunderstanding:
> with "warrior using AI techniques" I meant
> "warrior whose code implements some AI
> inside, like fuzzy logic".

No. I suspected that, which is why I recommended studying Redcode
first.

--
---
an evolved lifeform

--
Posted via a free Usenet account from http://www.teranews.com
 >> Stay informed about: AI warrior. 

SimFlyer wrote:
> On 26 Jul 2006 23:14:41 -0700, <cristian210273 DeleteThis @yahoo.it> wrote:
....
> Go hang out on the nano hill. I was fooling around with some numbers
> the other day and it seems that you could test every possible warrior
> for that hill in about 18 months (for my machine).

Hmm... I did some back of the envelope calculations

double D = 16*80*80*7*7*8;
Console.WriteLine(D);
// 40 million possible lines of redcode on Nano
Dn = D*D*D*D*D * 5;
// require 5 lines of code, 5 possible starting lines
Console.WriteLine(Dn);
// 5.2E+38 possible warriors on Nano
double Ds = 60*60*24*30*18;
Console.WriteLine(Ds + " secs per month");
// 46 million seconds in 18 months
D = Dn/Ds;
Console.WriteLine(D + " warriors per sec");
// 1.1E+31 warriors evaluated per second

and it looks to me like you'd need to test 10 raised to the power of 31
warriors every second.
What kind of computer do you have?

Dave Hillis
 >> Stay informed about: AI warrior. 

Cristian210273 wrote:
> >
> > The 5 instruction limit make that hill makes it a good place for
> > evolvers, including the human based variants.
> >
> > 'Neural Networks' - wrong hammer for this nail.
> >
> Thank you! I will investigate more deeply.....
> However there was a misunderstanding:
> with "warrior using AI techniques" I meant
> "warrior whose code implements some AI
> inside, like fuzzy logic".

You probably already know that Redcode is Turing complete, and also
that while it allows multiple processes, they aren't simultaneous-there
is no parallel architecture speedup. Whether one of these techniques
can be fast and compact enough for a warrior is the answer to a more
detailed question: what kind of neural net? on-line training?, etc.
My suggestion is to look at P-space warriors and see if you find
something that interests you.

Dave Hillis
 >> Stay informed about: AI warrior. 

>
> You probably already know that Redcode is Turing complete,
>
Dave, please help me! What does this mean?
(Sorry if this question sounds silly...)

> Whether one of these techniques
> can be fast and compact enough for a warrior is the answer to a more
> detailed question: what kind of neural net? on-line training?, etc.
> My suggestion is to look at P-space warriors and see if you find
> something that interests you.
>
Yes, my idea was to find an AI tecnique compact enough to be
coded in a warrior (I was thinking at fuzzy logic: a membership
function should not be so heavy in terms of number of lines
of code....). I will take a lokk at P-space: can you point me to
a reference?
Thanks!
 >> Stay informed about: AI warrior. 

First, since you posted the question at all, it means you would
probably enjoy playing Corewars. Even if what you have in mind didn't
lead to any successful warriors, similar things would.

Cristian210273 wrote:
> >
> > You probably already know that Redcode is Turing complete,
> >
> Dave, please help me! What does this mean?

I'll say it a different way. Warriors are written in an
assembly-like language called 'Redcode.' It is a complete language, so
pretty much anything you can write in a "normal" computer language, you
can write in Redcode (somehow...).

> (Sorry if this question sounds silly...)

You are not asking silly questions. For general questions, the
best way to answer is the way SimFlyer did: encourage reading the docs.
More specific questions tend to get more satisfying answers.

> > Whether one of these techniques
> > can be fast and compact enough for a warrior is the answer to a more
> > detailed question: what kind of neural net? on-line training?, etc.
> > My suggestion is to look at P-space warriors and see if you find
> > something that interests you.
> >
> Yes, my idea was to find an AI tecnique compact enough to be
> coded in a warrior (I was thinking at fuzzy logic: a membership
> function should not be so heavy in terms of number of lines
> of code....). I will take a lokk at P-space: can you point me to
> a reference?
> Thanks!

A good way to start is to go to http://koth.org/ look there,
scroll down to the Corewars Web Ring and look at the many guides you
will find on the various sites.
You'll see info on P-Space. It isn't the first thing to learn but
I will say a little about it. When 2 warriors compete, they fight a
sequence of battles. Each battle starts with the warriors in a clean
core (memory space) which they then mess up while trying to kill each
other.
P-Space is persistent memory each warrior has that only they can
read or write into, so they can pass information to themselves from one
battle to the next. One way to use this memory is a p-switcher. For
instance, a warrior might play one strategy during one battle, if it
wins it tries it again, if it loses it tries something else next
battle. More elaborate strategies are often encoded into finite state
tables. Your fuzzy rules might fit in for something like this.
I said that the other warrior can't write into your warrior's
P-Space. But it might trick your own warrior into writing harmful
information into it. That's called "brain washing."
Happy reading!

Dave Hillis
 >> Stay informed about: AI warrior. 

Some very basic AI can be studied under p-space.

AI learning. in term of GA, is not possible on the hill...not much time
to analyse anything. However, to do AI , GA , GP etc is to do it off
the hill against benchmark warriors. Even this tasks is VERY
challenging for 8000-core hill at the moment.

Fuzzy logic also not very useful because you never know what type of
warriors you are fighting. Even with known warriors..e.g. latest
silks/stone, it is quite a task to set the best strategy that worth the
time spend.

May be you can start with simple, old warriors eg Rave, Mice and Dwarf
and try to set something fuzzy versus them. May be you will discover
something if you are lucky

Zul Nadzri
 >> Stay informed about: AI warrior. 

>
> Fuzzy logic also not very useful because you never know what type of
> warriors you are fighting. Even with known warriors..e.g. latest
> silks/stone, it is quite a task to set the best strategy that worth the
> time spend.
>
Uhmmmm...... are you sure?
As I have already underlined, I'M A NEWBIE, so I could
tell a rubbish.......
You don't know which warrior you're fighting, but.......
this morning I read some tutorials that pointed out that
some basics tecniques are common to many warriors,
like vampire, scanning, imp and replicator.....
using fuzzy logic you probably can identify what kind
of attack is using your opponent, i.e. by analyzing the
memory you can obtain a fuzzy measure of each known
(I repeat.... KNOWN) tecnique.....
Am I wrong something in my reasoning?
Thanks!
 >> Stay informed about: AI warrior. 

In my opinion...

Some of the many challenges to make fuzzy logic a success:
1) The warrior must allocate a fuzzy logic bank in the body thus making
the size bigger:
a) Bigger warrior is easier to locate by the opponent
b) May be bigger than any pspace engine around

2) Need to allocate time to analyse the core and process the data:
a) May be some disadvantages against quick opponent

3) Need to find something fast in the core:
a) Some warriors don't have trails while others leaving big junk before
booting
b) Need to limit searches before deciding to abandon 'empty' search

4) Need to find a relevant discovery:
a) Decoy makers exist (I like them
b) Need an efficient way to differentiate what has been found (e.g.
what do you want to do if you found e.g. "spl #1, #1" or "add #3, #1"
or "mov @4, {7"

5) Need to store memory:
a) P-space definitely required
b) Beware of brainwashers
c) Also beware of p-warrior, it will make the fight fuzzier

6) Counterattack
a) Once the opponent identified quite correctly, how to deal with it?
Looks like a possible a p-warrior component required...

If these can be addressed efficiently, you will produce a worthy
warrior addition to the pool

Zul Nadzri
 >> Stay informed about: AI warrior. 

Cristian210273 wrote:
> Uhmmmm...... are you sure?
> As I have already underlined, I'M A NEWBIE, so I could
> tell a rubbish.......
> You don't know which warrior you're fighting, but.......
> this morning I read some tutorials that pointed out that
> some basics tecniques are common to many warriors,
> like vampire, scanning, imp and replicator.....
> using fuzzy logic you probably can identify what kind
> of attack is using your opponent, i.e. by analyzing the
> memory you can obtain a fuzzy measure of each known
> (I repeat.... KNOWN) tecnique.....
> Am I wrong something in my reasoning?
> Thanks!

You probably can analyse the warrior type from looking
at the core, but the problem is that you will spend so
much time in analysis and so little time attacking that
the dumb warriors will beat you to a pulp.

A warrior that tried the analysis approach with some
success in the past was Killer Instinct by Anders Ivner.
Sadly, QScans, and more importantly, lack of viable
vampires on the hill has obsoleted the techniques used.

Ian
 >> Stay informed about: AI warrior.