Most of my research for the last 30 years has been
around the idea of building systems incrementally using
a learning or knowledge acquisition stragegy known as
Ripple-Down Rules.
Contact Professor
Paul
Compton School
of
Computer Science and Engineering The
University of New South Wales Sydney
2052
Australia
mob 61 425375279
email
p.compton@unsw.edu.au Rm
303 Computer Science Building (map ref k17)
Ripple-Down Rules (RDR) is a strategy of bulding
systems incrementally while they are already in use.
When a system does not deal with a case or situation
correctly a change is made in such a way that the previous
competence of the system is not degraded. The change
is made simply and rapidly and the difficulty of making a
change should not increase as the system develops.
RDR can be categorised as a type of apprentice
learning
Various industrial
RDR systems have been developed for a range of
applications. There have have been research proofs
for a wide range of applications including: various types
of classification problem, configuration or parameter
tuning, text-processing, conversational agents, image
processing, heuristic search, tuning genetic algorithms
and multi-agent environments. There have also been
machine-learning versions of RDR.
A number of researchers across various universities have
been involved in this research. It
seems current research is aimed at:
- extending the range of possible applications of RDR,
particular with text processing and conversational agents
- further extensions to RDR algorithms
- ways of combining machine learning and RDR, in
particular RDR systems that can recognise when a case is
outside their competence
Starter Papers
These following material should provide some introduction to RDR,
while the book[1] attempts to cover all known RDR. There a
numerous other RDR papers by other authors as well as my coworkers
and myself. In particular in 2009 Debbie Richards published a
paper covering 20 years of RDR research at this stage
[1]Compton,
P. and Kang B.H (2021) Ripple-Down Rules: the Alternative to
Machine Learning, CRC Press (Taylor and Francis)
The central part of this book is three chapters providing detailed
worked examples for various types of RDR. The reason for these
chapters is that although RDR are very simple, people often assume
they must be doing something more complex. The examples were
developed using software available here.
This software is written in VBA and Excel, to make it easy for users
to play with RDR using their own data; it is fully featured, but not
appropriate for industrial application. The book also
discusses machine learning because of current interest in
learning. Machine learning algorithms are magical, but often
the data is simply not good enough for the magic to work. RDR
provides an alternative approach that avoids the need for a large
well-curated data set.
[2]Compton,
P.
(2013). "Situated cognition and knowledge acquisition research." International
Journal of Human-Computer Studies71: 184-190.
This was one of 11 invited papers
for a special issue of the journal to commemorate 25 years of
knowledge acquisition research. It summarises why RDR is
different from other knowledge acquisition strategies, and
presents some data on RDR performance.
[3]Compton,
P., L. Peters, T. Lavers and Y. Kim (2011). “Experience
with long-term knowledge acquisition.”Proceedings of
the Sixth International Conference on Knowledge Capture, KCAP
2011, Banff, ACM. 49-56
This paper presents more
detailed data on some industry RDR systems. A version of this
paper with minor corrections is available on the PKS
web site.
[4]
Compton, P. and R.
Jansen (1990). "A Philosophical Basis for Knowledge
Acquisition." Knowledge Acquisition2: 241-257.
This was the second RDR paper, but is the key paper for establishing
the philosophical basis for an RDR approach. This paper first appeared in EKAW
89 PDF
[5]Edwards,
G., P. Compton, R. Malor, A. Srinivasan and L.
Lazarus (1993). "PEIRS: a pathologist maintained expert system
for the interpretation of chemical pathology reports." Pathology25: 27-34.
This
was the first RDR system in industrial (clinical use). Although a
large system with 2000 rules, it was limited by using single
classification RDR
[6]Ho,
V. H., P. Compton, B. Benatallah, J. Vayssière, L. Menzel and H. Vogler (2009). “An
Incremental Knowledge Acquisition Method for Improving Duplicate
Invoice Detection.” Proceedings of the 25th IEEE
International Conference on Data Engineering, ICDE 2009,
Shanghai, IEEE. 1415-1418
This is one of a number of papers, where an RDR system is used to
correct the errors of another system, functioning as a sort of
wrapper. If the underlying system makes an error an RDR rule
is written. The conclusion from the underlying system is the
default conclusion, but is overridden if the RDR system provides a
different conclusion for the case. The underlying system might
have been developed by machine learning but potentially can be any
system that processes a case
