The other way to satisfy a goal is to subgoal on obtaining LCW,
and then expand the universal base. However, since LCW goals can also
get clobbered by subgoal interactions, XII has to ensure that actions
introduced for sibling goals don't cause the agent to lose LCW. For
example, given the goal of finding the lengths all files in /papers,
the agent might execute wc *. But if it then compresses a
file in /papers, it no longer has LCW on all the lengths.
To avoid these interactions, we use LCW links which
are like standard causal links except that they are
labeled with a conjunctive LCW sentence. Since LCW(P()
Q(
)) asserts knowledge of P and Q over all the members of the set
, an LCW link is threatened when
information about a member of the set is possibly lost or a new member, for
which the required information may be unknown, is possibly added to the
set. We refer to these two cases as information loss and domain
growth, respectively, and discuss them at length below. Like threats
to ordinary causal links, threats to LCW links can be handled using demotion, promotion, and confrontation. In addition,
threats due to information-loss can be resolved with a new technique
called shrinking, while domain-growth threats can be defused by a
method called enlarging.