Page 241 - Deep Learning

224 Adaptation

Figure 7.2. A schematic rule genealogy. The top rule recommends its Action when
condition C0 is satisfied. In Learning Event 1, two new versions are created by adding
conditions C1.1 and C1.2. In Learning Event 2, the first of these new, more constrained
variants is in turn specialized by adding conditions C2.1 and C2.2; and so on.

If the root rule is entered into the conflict set and if it wins against alterna-
tive rules, it will control action in that operating cycle. This is counterintuitive.
The root rule was revised because it generated errors. How can the cognitive
architecture eliminate errors if an erroneous rule remains in memory and
retains the possibility of grabbing control of action? The answer requires a
more careful analysis of what happens to the domain of application of a rule in
the course of specialization.
If a rule R that recommends action A applies in some set {S} of situations,
and if that rule is specialized, {S} is functionally speaking split into three sub-
sets: one subset in which the descendant Rʹ applies, a subset in which the other
descendant Rʹʹ applies and a third subset that contains all other situations. We
know that some of the situations in the third subset are those in which A gen-
erates an error, but it might also contain some situations in which it is as yet
undecided whether A is correct or erroneous. To keep the parent rule R in
memory after it has been revised is to retain the option of doing A in the lat-
ter class of situations. If A once again generates an error, it will be specialized