| RippleDown™ Rules |
| |
| Summary |
| |
| RippleDown™ Rules technology is a radically different way for information systems to learn professional
and business expertise so that this can be used to improve the quality and cost-effectiveness of products and services. |
| |
| The core strategy which differentiates RippleDown™ from other knowledge management/ business rule/ expert
system technologies, is that acquisition of new knowledge is incorporated into the human trainer’s other activities.
It has almost no impact on workflow and workload – except that workload goes down as the RippleDown™
system gradually takes on more of its trainer’s duties. Knowledge can be endlessly changed and evolved
as business requirements change – with the cost of adding each new chunk of knowledge no greater than the first. |
| |
| Background and Problem |
| |
| The early value provided by computing technology was in high-speed calculation, taking over from the abacus and
mental arithmetic. This was gradually merged with the storage and retrieval of data records, taking over
from the filing cabinet and clerk. We are all well aware of the most recent change: the anarchic global
supply and availability of information over the World Wide Web – a global bulletin board, where you need special
assistance to try and find notices of interest. |
| |
| Another line of development has been the possibility of computers taking over, or at least assisting in human expert
decision-making. This was also expected to transform the value provided by computing. However,
the promise of knowledge-based or expert systems was greatly oversold, and led to the very sceptical “Artificial
Intelligence winter” of the 80’s. |
| |
| The key insight that lay behind expert systems was the realisation that experts such as lawyers or doctors, might
be very intelligent people – but what made them valuable was not really their intelligence and reasoning ability,
but the fact that they knew a whole lot about their specialist area. This lead to the idea of developing
computer programs where the inference engine or reasoner was separate from the knowledge the system had to reason
about. There were huge advantages in this. For example in a system that expressed knowledge in rules,
the user simply had to enter a rule; it was the inference engine that would work out when to use this rule, how
to decide whether it was the new rule or a previous rule that should be used. It was thought that to
build one of these expert systems one would simply get the expert to give a whole list of rules of what to do in
various circumstances. |
| |
| The reality fell far short of the promise: a separate class of programmer, the knowledge engineer, emerged
to try to deal with the difficulty of getting sufficient and adequate rules from the experts. The phrase
the “knowledge engineering bottleneck” was coined to express the difficulty of building these systems[1] The essential problem is that the knowledge engineer
needs the expert or business manager to tell them every detail about the domain. Following Grosof; it is
not sufficient for the business manage to provide a rule: |
| |
| If buyer returns the purchased good for any reason, within 30 days, then the purchase amount, minus a
10% restocking fee, will be refunded |
| |
| and then another |
| |
| If buyer returns the purchased good because it is defective, within 1 year, then the full purchase amount
will be refunded |
| |
| The business manager must also supply a rule like: |
| |
| If both Rules A and B apply, then Rule B ‘wins’, i.e. the full purchase amount will be refunded. |
| |
| because the inference engine by itself will not be able to work out which rule to apply. It is
extremely difficult to get experts to provide all this knowledge[2]
In medicine it is almost impossible to get a clinician to explain all the combinations of signs and symptoms that
may occur in a disease – and to also indicate all the circumstances when these signs and symptoms are not symptomatic
of the disease. |
| |
| Other solutions |
| |
| Early attempts to solve the knowledge-engineering bottleneck looked at whole range of better ways to get experts
to really express how they were thinking. This made little real difference: there were always some bits of
knowledge left out and to go back and try to understand what had been done before, to find what to change remained
a major challenge. One hears comments that changing or adding 2-5 rules per day to a large knowledge base
is the industry norm. |
| |
| For the last decade the major research focus has been on knowledge-level modelling. This is an attempt to
provide a systematic and abstract framework firstly for understanding the different types of expert tasks that
are possible e.g. diagnosis identifies the faults that cause the symptoms, with medicine as the most obvious example;
resource allocation covers tasks like assigning offices in an organization. This approach then provides
an analysis of the types of methods and their interrelationships that can be used in solving such problems.
Finally the there are frameworks for how knowledge about various domains can be structured and organised. |
| |
| Any student of problem-solving and reasoning should study this work; it has been a major scientific advance.
If a knowledge engineer is trained in this type of analysis, there is no doubt they will have a better chance of
understanding some new problem type and succeeding in building and expert system. However, the expert will
still be expected to provide all relevant knowledge and there will still be the problems of incorporating other
knowledge when the need for this is realised. |
| |
| It is worth noting that developers who are unfamiliar with the history of expert systems and the consequent body
of research on acquiring and organising knowledge, keep developing the same false expectations that since individual
rules are so easy to add for small knowledge-bases; large rule-based systems are easy to build. The database
community has recently re-discovered the advantages that will accrue if business rules are written separately from
the reasoning engine, rather than low level SQL code (e.g. Date, C. J. What, not how: the business rules approach
to application development. Reading, Mass., Addison Wesley 2000). However, probably because they see
the value and role of business rules as quite different from the earlier push for expert systems, they are yet
to realise the advantages they have discovered have their own further problems. |
| |
| Finally, it should be noted that there is an on-going research program in how to learn what people do without them
having to explain what they do. A wide range of machine-learning techniques including neural networks have been
developed[3]. Although these are very powerful
they still only apply to some limited types of problem and require large amounts of well-worked-up data and may
require considerable tuning and adjustment of the learning software. They are finding most use in data
mining, i.e. in searching for patterns in large data bases. |
| |
| The RippleDown™ approach |
| |
| The central feature of RippleDown™ is that it steps aside from the goal of systematically understanding
organising and assembling all the knowledge in a domain. It sees this as an unachievable dream, the inheritor
of the Platonic belief that somewhere in the universe there exists archetypal, canonical knowledge[4]. |
| |
| In the RippleDown™ approach there is no knowledge engineer who is expected to organise the domain knowledge.
Nor is the expert expected to provide some sort of integrated view of all the knowledge. The expert does
not need to know how each new piece of knowledge is incorporated into the knowledge base – in just the same way
as we do not need to know how are minds store what we know in order to assimilate some new piece of information.
RippleDown™ achieves this by having a refinement structure which automatically locates rules in the knowledge
base in such a way that they will only be used in the same context in which they were provided[5]. |
| |
| The complementary feature of RippleDown™is that the system also assess whether the new knowledge provided
conflicts in anyway with previous knowledge. It does this by retrieving past cases to which the new knowledge
might apply and asking the expert whether this is the case. This prompts the expert to identify extra discriminating
features. Again this is analogous to how human trainees interact with their teacher, particularly in on the
on the job training. |
| |
| The result of the RippleDown™ approach is that knowledge bases can be built gradually over time.
The original goal of being able to simply add a rule with no concern for the other rules already in the system
is at last achieved. The task of adding a new piece of knowledge always remains the same, regardless of the
size of the knowledge base. Since it generally takes about one minute, the development of the knowledge base
can be incorporated into the expert’s normal duties, with the system gradually evolving and changing as required. |
| |
| Of course nothing is ever quite as simple as it seems. RippleDown™ still requires some initial
knowledge/software engineering in order to link to the information system providing the data to be reasoned about
and to identify particular concepts the expert wishes to use. However, this is a comparatively small
one-off problem. Next, although RippleDown™ can be used to address a very wide range of expert
system/ business rule/ knowledge management problems, it can’t yet address all of these problems. Our partners
at the University of New South Wales have a long-term research project in extending the technology. They
have also carried out research analysing the performance characteristics of the technology to explain its remarkable
success. (The results of their research can be found here.) |
| |
| Clarification of Ripple Down Rule features |
|
Superficially, Ripple Down Rules would appear to be similar to other Artificial Intelligence technologies, in
particular, to Case Based Reasoning technololgies.
However, Ripple Down Rules are based on different assumptions and are used quite differently, despite these
apparent similarities.
Contact PKS to receive further information on the following topics:
- the unique features of Ripple Down Rules
- how these features distinguish Ripple Down Rules from other technologies
- the benefits of Ripple Down Rules over conventional technologies.
|
