[Freeciv-Dev] Re: [Wishlist] Attitude manager (agent and AI)

[Raimar Falke <hawk@xxxxxxxxxxxxxxxxxxxxxxx>]

On Mon, Mar 04, 2002 at 02:35:21PM -0500, Jason Short wrote:
> Gregory Berkolaiko wrote:
> > On Mon, 4 Mar 2002, Raimar Falke wrote:
> > 
> > 
> >>On Mon, Mar 04, 2002 at 05:42:25PM +0000, Gregory Berkolaiko wrote:
> >>
> >>>>>So I would suggest the following extension to the CMA: once the
> >>>>>target multipliers are set, you should be able to evaluate the
> >>>>>effect of building a happiness-related improvement.  It would output
> >>>>>(for user) the increase in surpluses you'd be able to get (maybe
> >>>>>minus the cost of the upkeep).  And for AI you just need to combine
> >>>>>the values of increases multiplied by their _WEIGHTINGs
> >>>>>
> >>>>I don't know how this relates to CMA and I don't know what target
> >>>>multipliers you mean.
> >>>>
> >>>Now let's consider evaluating impact of a Cathedral.  The algorithm number
> >>>2 above will not give us anything useful (unless the whole city changes
> >>>it's happiness status).
> >>>
> >>I don't see a big problem in creating a happiness surplus value based
> >>on the ppl_* arrays.
> >>
> > 
> > Well it would be useless, that's the problem.  The value of happiness is 
> > to allow more people to work.  To estimate the effect of them working 
> > you'd need to estimate F/S/T obtained through their (best) placement.  AI 
> > can do it poorly for itself, CMA can do it perfectly for the human.
> 
> I think you guys are missing (or at least ignoring) something here.
> 
> Yes, the value of happiness is to allow more people to work.  A simple 
> CMA should be able to evaluate building a cathedral, building a factory, 
>   or changing the global luxuries rate.
> 
> But where this model falls short again is that it only evaluates the 
> current situation.  For example, it is easy to deduce that with no 
> miltary units in the field a police station will not provide any 
> benefit.  It is less easy to determine that we want a police station 
> before churning out units.
> 
> As a more extreme example, you've said that the value of happiness is to 
> allow more people to work.  Well I say the value of _food_ is to allow 
> more people to work.  If a human is using the CMA, it is enough that the 
> CMA tell the human what extra productivity they will get out of a 
> cathedral.  But for an AI to use this, we need to know more to choose 
> our goals.  Spending 30 turns to build a cathedral may give us 2 extra 
> workers in 30 turns.  Or we may divert more production into food and 
> have 1 extra worker in 15 turns instead of 30, but at the cost of 
> delaying the cathedral by 15 turns.  In this case, I think a CMA that 
> simply does a linear optimization will not be very helpful - the 
> higher-level agent will need to know more.  Alternately, it might be 
> possible for the CMA to itself determine the usefulness of food - 
> perhaps given an extra constraint of time weighting.
> 
> This problem becomes even harder when considering, say, founding a new 
> city.  40 units of production + 1 citizen (which we can think of in 
> terms of food, perhaps) gives us 2 citizens.  It is basically a straight 
> conversion from production to food, but with a slight overhead because 
> we have to move the settler.  Then we have to consider how productive 
> the citizens are (which will include our increased happiness at the 
> current city after reducing our population).  Finally, we need to 
> consider that, since granary size increases as the city increases, 
> founding a new city will effectively reduce the cost for future growth.
> 
> I don't think the agents-AI is at a point yet where we should begin 
> thinking about implementing these things.  Things should go forward in a 
> straightforward manner until it hits a wall.  At that point hopefully 
> we'll know enough to work on the _real_ problem.

In general ack. Two points: the CMA will not be extended this
much. Probably supporting virtual cities is needed but not the other
stuff mentioned.

Second point: these model are IMHO too fine grained. You have larger
fluctuations if an enemy is near and your settler needs to use a safer
path or an enemy is near and the settler have to completely abort its
mission. If we are at the point were we can arm our models in such a
way we could make tests to see if this really gives a better AI.

        Raimar

-- 
 email: rf13@xxxxxxxxxxxxxxxxx
  "The primary purpose of the DATA statement is to give names to
   constants; instead of referring to pi as 3.141592653589793 at every
   appearance, the variable PI can be given that value with a DATA
   statement and used instead of the longer form of the constant. This
   also simplifies modifying the program, should the value of pi
   change."
    -- FORTRAN manual for Xerox Computers