Re: [Rails-devel] Route calculation for huge runs

[Stefan F. <ste...@we...>]

On Wednesday, August 10, 2011 07:37:29 am John A. Tamplin wrote:
> On Wed, Aug 10, 2011 at 1:19 AM, brett lentz <bre...@gm...> wrote:
> > Personally, if y'all want to pitch in and use Rails to try out
> > different algorithms and see who can come up with the best one, I
> > think it would be fantastic. It may even be a good foundation for
> > supporting AI, which has also been a long-time dream. ;-)
> 
> AI is going to have to be able to very quickly (like in ms) evaluate
> possible runs in various scenarios, because it is going to have to evaluate
> many options and hopefully a few moves ahead.  For that, it doesn't have to
> be accurate, just reasonably close -- but it absolutely has to be
> blindingly fast.
> 
> For regular route calculation, I would hope it would be better than Simtex
> -- I frequently found incorrect runs.  If we get people relying on it and
> then they find it produces incorrect answers, then that will be a problem.
>  I would prefer to come up with some heuristic for deciding when the
> exponential approach will take too long, and then do some approximation and
> label it clearly as such.

The current setup is strongly optimized for the regular calculation (the graph 
is always created from scratch and there is no easy way to simply add or 
change a tile once it is created). 

This was somehow different from Alex' approach one year ago (what happened to 
his prototype?) who tried to keep the door wide open for AI.

I agree with John that a separate approach to route calculation for AI is 
optimal. For AI you are mainly interested in finding revenue increases by 
changing the graph in several steps, thus the strategy space is again 
exponentially larger. And then you have to rely on all kind of approximations 
and "non-correct" tree pruning. And you are allowed to do so for AI.

Stefan