Complete.Org: Mailing Lists: Archives: freeciv-ai: May 2002:
[freeciv-ai] Re: long-term ai goals
Home

[freeciv-ai] Re: long-term ai goals

[Top] [All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index] [Thread Index]
To: "Ross W. Wetmore" <rwetmore@xxxxxxxxxxxx>
Cc: Per I Mathisen <per@xxxxxxxxxxx>, freeciv-ai@xxxxxxxxxxx
Subject: [freeciv-ai] Re: long-term ai goals
From: Raimar Falke <rf13@xxxxxxxxxxxxxxxxx>
Date: Sat, 25 May 2002 18:49:05 +0200

On Sat, May 25, 2002 at 11:39:44AM -0400, Ross W. Wetmore wrote:
> At 12:19 PM 02/05/25 +0200, Raimar Falke wrote:
> >On Sat, May 25, 2002 at 12:56:40AM -0400, Ross W. Wetmore wrote:
> >> Note there is a difference between a Freeciv AI designed to provide a
> >> challenge to a standalone player and 
> >
> >> the client agents project to do extensive but menial computations
> >> that assist human players. The latter are a completely different
> >> type of beast whose intelligence capabilities are (hopefully)
> >> provided from an external source.
> >
> >Yes the agents doesn't have buildin intelligence. Or more formal
> >speaking: the user has to provide the utility function which says how
> >good a given result state should be viewed.
> >
> >At point up in the hierarchy there will/should/may be some
> >intelligence. I have no ideas how this turns out but I'm pretty sure
> >that this level is above the management of a single city. The
> >parameter to manage a single city are only a few.
> >
> >     Raimar
> >-- 
> 
> Good. We are on the same page here (and thus there are significant 
> feelings of relief). 
> 
> I agree that once the lowest level of mechanical assistants are 
> available, a management layer may start to provide more strategic 
> direction. But I think this is significantly far out. Freestanding
> analysis and control at the very top is even further out. These are 
> probably further out than necessary because of the decisions to 
> work on detailed individual agents first rather than a dynamic 
> framework for connecting agents with rough Syela-like parameterized 
> forms as initial offerings.
> 
> The code to do memory is trivial and by itself useless. This will 
> require both sophisticated analysis and interpolation/extrapolation
> to produce data in which old and now inaccurate data is now fuzzed
> in some useful way, and data that does not exist is introduced in
> correspondingly fuzzy ways. The latter is the real problem that FoW
> as it stands exacerbates to the point of killing the AI if it fully
> respects the rules. (An AI that only knows about its immediate city
> environs cannot even compute the weight for a city founder under
> strict FoW, has few trigger events to start building patrol units, 
> and thus will probably think that massive improvement construction 
> is the only way to nirvana - note this is a correct approach if the
> map really does only consist of the known city environs :-).

Another example I came up lately: both form of amortize (current CVS
and SMA) don't handle the fact that for the short action you have a
free settler in a shorter time. So for the finite time approach you
have something like:
 
 total benefit of action = sum over all surplus from delay till
 turns_to_consider + surplus that the settler can produce in
 (turns_to_consider-delay) turns

The production of a settler per turn (which is something <1 food,
shield and trade) can be calculated from the history of all
settlers. Well not at game start but latter.

> Since this sort of data munging can be easily simulated by the server
> there is no reason not to introduce ways for the AI or a human with
> handicapped intuition abilities to have a "server agent" or data
> distribution system provide this function during the next several 
> decades while AI research and technology is being developed for a 
> standalone version.

But the data has to be at the client. If we want to do client side AI
everything which is on the server will be slow. So quering the server
for every piece of data will not work.

        Raimar

-- 
 email: rf13@xxxxxxxxxxxxxxxxx
 Make a software that is foolproof, and only fools will want to use it.


[Prev in Thread] Current Thread [Next in Thread]