[Freeciv-Dev] Re: CSS was: Re: auto settlers rework

[Cyrill von Planta <pla@xxxxxxxxxxxxxxxx>]

Raimar Falke wrote:
> 
> On Tue, Jan 22, 2002 at 07:04:38PM +0100, Christian Knoke wrote:
> >

> > > > > The only way AFAIK is to increase the time horizon of the
> > > > > calculation. The current server auto-settler code is based on 24 turns
> > > > > (the famous MORT constant). This means that the action which would
> > > > > yield the most benefit in the next 24 turns is considered.
> >
> 
> > The time horizon is not linear. Things I can achieve (points I can get)
> > earlier in the game are more valueable than things I get later.

> 
> As little interaction as possible. I would prefer to add another
> layer/agent on top of these which changed the weights according to
> some algo. But I agree it should be possible to add extra city
> depending weights.

just an idea:
this algo could be temporal difference learning. this algorithm doesn't
require a completely correct or complete (freeciv)world model to learn
"good" weights.  this means it could probably do meaningfull actions
without too much computation.
it also incorporates the concept of a discounted sum of future rewards.

cyrill