[Coin-discuss] two phase algorithm with algo vars in BCP continued

[jvielma at entelchile.net]

> Hello Juan Pablo and everyone else,
> 
> I try to reply to your comments in a single email.  
> 
> First of all, the 2 phase algorithm is tested, but barely... It just
> doesn't come up that often. My comments on your questions are below,
> but first a general comment. It seems that the multi-phase code
> doesn't work properly at the moment. I'm certain that it can be
> cleaned up reasonably quickly, it's just that I don't have time to
> work on it now. If you are willing to dig into it, I'd be more than
> happy to answer your questions, help you along the way and finally
> incorporate the fixes into BCP.

Ok, I'll do that. That will give me a chance to get more familiar with the complete BCP code.
 
> 
> As I explained above, pricing algo vars in the root (while not
> impossible) may gain very little, especially if the algorithm the user
> has for generating the algo vars cannot take advantage of information
> of the kind "no need to consider this sort of variables". Still, I'd
> be very interested in trying out things.

I had two reasons to price algo. vars in the root:

1) Fix variables by reduced costs: as you said, for algo vars this is of no use as I should just not generate those vars. This could be acomplished by generating algo vars under some rc bound insted of just  generating ones with negative RC

2) To get a LP bound: In my particular problem I don't have ( yet ) a good bound for a not fully priced LP, so I need to do full pricing for the root LP to be able to validate some GAP.

I think 2) might be a good reason to alow the possibility of full pricing on the root at some moment.