Re: [Help-glpk] glpk 4.59.2 non-official test release

[Andrew Makhorin]

> >> Another thing that I noticed (but is probably normal behaviour) is
> >> that --proxy takes some time to find a solution when compiled with
> >> optimisations and x87 floating point, where using only --bestp a first
> >> solution is found quickly with both sse and x87 floating point.
> >
> > To find a first integer feasible solution it is better to use --bestp
> > rather than --bestb (the latter is default option). The best projection
> > heuristic (--bestp) selects a subproblem whose solution to lp relaxation
> > is close to an integral point while the best bound heuristic (--bestb)
> > attempts to minimize the size of the search tree.
> 
> What I was trying to say is that the first stage of proximity search
> (as implemented) is using the solver with best projection search to
> find an initial solution, 

No, currently the proxy routine does not change the selection heuristic
option using the default strategy which is the best bound search, not
the best projection one.

> so I was (naively) expecting similar
> performance when using --bestp on the command line. However the
> proximity search one takes an order of magnitude more iterations for
> the first solution. I understand that the search is probably just
> taking a different path, but maybe it offers a hint.