Hierarchical Case-Based Reasoning Integrating Case

Hierarchical Case-Based Reasoning Integrating
Case-Based and Decompositional ProblemSolving Techniques for Plant-Control Software
Design
Authors:
Barry Smyth, Mark T. Keane,
Padraig Cunningham
What is Case Based Reasoning (CBR)
CBR System uses case bases of
previously solved problems to solve new
problems
The solution is modified to fit the new
problem
What is Hierarchical Case Based
Reasoning (HCBR)
HCBR is a technique created by the authors for
multiple-case (more complex) reuse
Uses multiple levels of solution abstraction
Abstract case is like a decomposition template
Parts of previous solutions, stored as individual
cases, can be reused and recombined to solve
some subproblem
What is Hierarchical Case Based
Reasoning (HCBR)
A decomposition template’s abstract solution
is used to break up a complex problem into
subproblems
These subproblems are solved from the
solution code segments that are produced
from the reuse of concrete cases
The integration of these code segments is
guided by the solution structure of the
abstract cases
What is Automated Programming
Starts with a specification of some task
Ends with a program that is the solution
to this task
Abstraction Levels
Abstract plans outline solutions
Abstract plans are refined by replacing
abstract operators with collections of
more detailed operators
Complete plan will contain only
primitive operators
Different Cases
Abstract Cases
Solutions correspond to high-level plans for
particular problems
High-Level program designs
Concrete Cases
Solutions correspond to actual programs
Combining Cases
Abstract Cases
Solution Part:
Contains abstract operators that
correspond to high-level actions
Abstract operators also act as subproblem
specifications
Description Part:
Similar to description part of a concrete
case
Drawn from the abstract task hierarchy
Complete Set of Abstract Tasks
Concrete Cases
Solution Part:
Sequential function chart made up of
primitive operations
Description Part:
Contains a set of features that relate to the
case solution
In déjà vu these descriptions are task
oriented
Complete Set of Concrete Tasks
The HCBR Process Model
Decompose a specification into right set
of subproblems
Solved separately
Individual solutions can be recombined to
produce a suitable overall solution
The Main HCBR Algorithm
Retrieval and Adaptation
Retrieval
The target specification features are matched
against the description features of cases in the
case base and a measure of similarity is
computed
Result is a ranking of cases according to their
similarity to the target
Adaptation
Ensures that the retrieved case is the easiest of
those available to adapt, not just the case that is
most similar
Decomposition
Integration
Each HCBR cycle generates one new solution
component
Integration adds this component to the
evolving solution
Solution is an abstraction hierarchy and each
new solution component is added as a leaf node
Integration
Learning
New Solved Problems Learned By:
Packaging their specifications and solutions
together as new cases
Adding these new cases to the case base
Parts of the problem can be learned as
separate cases at various levels of
abstraction
Condition tests whether or not the new case
is novel enough to warrant addition
Results
Benefits of HCBR
Repeated solution segments are not redundantly stored
within multiple large cases
Instead they’re stored as single case instances that can be
easily reused
Avoids the problem of decomposing a problem for
which there are no suitable cases in the case base
Cases and the decompositional knowledge are fully integrated
The collection of abstract cases in the case base is the
decomposition knowledge of the system (both use the same
sources of knowledge for all problem solving)
Improved problem solving coverage
Represents complex problems as collections of independently
reusable abstract and concrete cases