[Freeciv-Dev] Re: [RFC][Patch] AI can fly v2 (fwd)

[Gregory Berkolaiko <Gregory.Berkolaiko@xxxxxxxxxxxx>]

On Fri, 12 Apr 2002 per@xxxxxxxxxxx wrote:

> On Fri, 12 Apr 2002, Gregory Berkolaiko wrote:
> > what do you propose as a measure of "good chance of winning".
> >
> > Examples:
> > 1. you put 1$ on red and it wins with prob 1/2 and you get back your 1$
> > plus 0.5$.  Is it good?  Do you play?
> > 2. you put 1$ on something that wins with prob 0.75 and you get back your
> > 1$ plus 0.3$.  It this good?
> > 3. you put 1$ on number 13 which wins with prob 1/32 and you get back you
> > 1$ plus 100$.  Is this any good?
> 
> None of those are good. We must also know if we can afford to lose $1.
> 
> I propose that we use the "minimax" principle: Assume worst case scenario
> - if we cannot afford to lose, don't do it. Some things to look for: Is
> the unit directly threatening a city? We cannot afford to lose the city,
> so attack no matter the odds. Are we outproducing the enemy? Take greater
> risks - we can afford more losses. (Or perhaps rather, somehow figure out
> if we are on the defensive, if we are not omniscient.) Are we on the
> defensive or being outproduced? Take less chances. Etc.

yeah, I wrote a couple of equations, I think we can work out a way to take 
into account this...  Of course we want to attack an enemy which is 
endangering us, but we should have this info first and it would take a 
real warmap to get it: a map which shows which tiles are trheatened by 
which units.

But in general, I deduced that the real measure of want is 
kill_desire / our_unit_cost or, expanding kill_desire,
        prob_to_win*(enemy_cost / our_cost) - prob_to_lose

This is a "want" independent of the cost scale you are on (in early stages 
units cost 10-20 shields, in later stages it's 100-200 shields).  It 
should be compared to the centered productivity ratio,
        enemy_total_production / our_total_production - 1

> Do it yourself. Send "subscribe freeciv-ai" in the body of a message to
> "listar@xxxxxxxxxxx".

done sent
now testing

> Ok, but after my updating frenzy yesterday, it only contains conndlg. Not
> much to test yet.


I know

G.