Information Technology Reference
In-Depth Information
The following takes concept of failure as an example to explain the
explanation and the formation of control rules. This is a problem of action plan
for domain background building block. For the learning of a certain failed plan
action, the yielded explanation is:
(OPERATOR-FAILS op goal node) if
(AND (MATCHES op (PICKUP
x
))
(MATCHES goal (HOLDING
))
(KNOWN node (NOT (ONTABLE
x
))))
Here the lowercase characters are variables. The above expression is the
result of generalizing the failed action by the formal knowledge in the domain
knowledge base and it means that if the current node is not “ONTABLE
x
x
” and
the current operator is “PICKUP
x
”, then the operator “PICKUP
x
” is a failed
operator. Thus the learned rejection rule from the failed concept is:
(REJECT OPERATOR (PICKUP
x
)) if
(AND (CURRENT-NODE node)
(CURRENT-GOAL node (HOLDING
x
))
(CANDIDATE-OPERATOR (PICKUP
x
))
(OPERATE-FAILS op goal node))
Here op= (PICKUP
x
), goal= (HOLDING
x
), node= (NOT (ONTABLE
x
)).
The rules mean that if the current node is (NOT (ONTABLE
x
)), current goal is
(HOLDING
) and the
operator fails under this node and this objective, then operator (PICKUP
x
), the candidate operator on the node is (PICKUP
x
x
)
should be refused.
By passing the control rule to four strategies which include node choose,
sub-objective, operator and a group method of constraints, the capability of the
problem solver can be dynamically improved. When dealing with similar
problem solving, selection rules are used to choose proper sub-sets, and then
rejection rules are taken to do the filter, at last preference rules are adopted to
search for the best heuristic choice to reduce the amount of search.
3. Knowledge representation
The knowledge base of PRODIGY is composed of domain level axiom and
construct level axiom. The former deals with domain rules and the latter contains
inference rules in problem solving. Both of them are represented in declarative
logical language to better extend the inference explicitly.
Though there is no distinct generalization in PRODIGY, its domain rules,
inference rules of problem solving in particular, have actually contain
generalization. They are not made up of primitive domain knowledge which
makes them generalizable. On the other hand, PRODIGY generates the
explanation by proof tree according to generalization, thereafter the results of the
