I've implemented the optimal cost partitioning in a way that hopefully takes the
first achievers into account in r3932 of the emil-new branch. Note that there
are many other differences between this branch and the trunk, e.g. to do with
action landmarks, and the cost assignment method cannot be used directly in the
trunk since the way in which it is called is different in that branch.

Sounds good

OK, we'll have to rewrite that code in the "emil-new" branch anyway for
conjunctive LM support, so I'll leave the trunk alone for you to decide what to
do about it.

Generally speaking, we made a bunch of changes to the landmarks code in that
branch, some of which are worth considering for the trunk (e.g. see issue74). We
cannot merge them easily since we also threw away several features, though.

How about discussing the landmark-related things after the ECAI deadline?

I can reproduce this, and the assumption about first achievers seems correct.
Using the old LP formulation yields a better estimate the uniform cost sharing.

PS: If there indeed is a problem in trunk, it's probably hidden by default
because action landmarks take care of this for the given Blocks instance. But it
should show up if you can disable action landmarks in some way.

In the emil-new branch, "olsp" gives a worse init h value (4) than "ols" (6) on
blocks problem 4-1. I cannot test this on trunk at the moment as I'd have to
mess with the Makefile to set up the LP solver there; Erez, can you reproduce this?

A possible explanation is that the new LP formulation doesn't distinguish
between first achievers and achievers, and hence it thinks that it can achieve
"clear(c)" and "on(c, a)" together even though it can't, and similarly for
"clear(d)" and "on(d, c)". (There exist operators that are *achievers* of both
clear(c) and on(c, a), but not *first* achievers.)