[Freeciv-Dev] Re: [RFC] Caching AI values

[Raimar Falke <hawk@xxxxxxxxxxxxxxxxxxxxxxx>]

On Tue, Mar 19, 2002 at 02:47:28PM +0000, Gregory Berkolaiko wrote:
> On Tue, 19 Mar 2002, Raimar Falke wrote:
> 
> > On Tue, Mar 19, 2002 at 01:39:34PM +0000, Gregory Berkolaiko wrote:
> > > While doing flying stuff, AI computes a few values which should be cached 
> > > (like the number of air targets around a city).  I can see 2 approaches 
> > > to 
> > > such caching:
> > > 
> > > 1. Compute these values at the start of the AI player's turn.
> > > 
> > > 2. Compute these values as needed and then record them into a struct 
> > > together with the time they were computed.
> > > 
> > > I favour the second approach greatly.  First, you don't have to trace the
> > > spagetti AI calls to see to it that the values are initialized before
> > > being accessed.  Second, you don't have to decide beforehand if you should
> > > waste CPU initializing them at all or it's too earely in the game to do
> > > that (doesn't make much sense to count the number of air targets before
> > > you discover Flight): in approach 2 they would be initialized as needed.
> > > 
> > > But to execute 2nd approach efficiently, one has to have the "time"  
> > > concept.  I remember Raimar and Co talking about game.turn variable or
> > > somethign like this.  My time has to be a bit more precise, I think: there
> > > are two AI movement phases per turn and "time" should distinguish
> > > between "turn 10 start-turn" and "turn 10 end-turn".
> > > 
> > > Please provide me with insightful remarks and great suggestion on these 
> > > matters.
> > 
> > IMHO the second approach favors the spagetti AI. I can image a
> > description like "if foo is called from bar with option X than the
> > value is updated else the value is set by foobar".
> 
> let's say "it adapts itself to the spagetti AI".
> unfortunately, some aspects of AI are inherently spagetti.
> 
> > IMHO the granularity of pre-turn, post-turn is too coarse. But this
> > depends on exactly what you try to measure.
> 
> Yes, it is a bit coarse but if you refine it further, these values will be 
> recalculated too often to give you any time saving.
> 
> > There is another solution possible:
> >  3. Update the value whenever this is necessary
> 
> You probably mean "whenever environment changes".  It is a very good 
> approach.  But again there are arguments against it:

> - too complicated: go trace all events which lead to "cnahge in 
> environment"

We already capsulate common events like got new tech, unit moved, city
build, ... IMHO update_research is a good example for this. It also
calculate values which are cached.

> - you might be doing all this in vain --- how do you know that these 
> values will ever be accessed?

Do it lazy.

> To implement 3 we would have to rewrite half of Freeciv to be event-driven 
> system.

I don't think so but this depends on the type of the events.

        Raimar

-- 
 email: rf13@xxxxxxxxxxxxxxxxx
 "Life is too short for reboots."