A trade-off between domain knowledge and problem-solving method


P. Compton1, Z. Ramadan2, P. Preston1, T. Le-Gia1, V. Chellen1, M. Mulholland3, D.B. Hibbert2, P.R. Haddad4, B. Kang5

1School of Computer Science and Engineering, University of New South Wales

Sydney 2052. compton@cse.unsw.edu.au (for correspondence)

2School of Chemistry, University of New South Wales

3School of Chemistry, University of Technology

4School of Chemistry University of Tasmania

5School of Computer Science, Hoseo University, Korea

ABSTRACT

The major focus of recent knowledge acquisition research has been on problem-solving methods (PSM). This paper present results where a PSM developed for classification has been extended to handle a configuration or parametric design task, designing ion chromatography methods in analytical chemistry. Surprisingly good results have been obtained seemingly because any knowledge that has been added to the knowledge base, has been added precisely to overcome any limitations of the PSM. These results suggest a trade-off between domain knowledge and the power of the PSM and that greater use of domain knowledge would facilitate re-use by allowing PSMs to be used for a broader range of tasks.

INTRODUCTION

The critical insight with the emergence of knowledge based systems (KBS), was the usefulness of knowledge as compared to pure search. This was formulated in ideas such as the "knowledge principle". This resulted in expert system shells which while specifically intended for the accumulation of knowledge as rules, also allowed developers great flexibility in the way they organised domain and problem solving knowledge (e.g. OPS5 (Brownston, Farrell et al. 1985) ). As fairly general purpose programming tools, these shells suffered from difficult maintenance problems like any other software (Bachant and McDermott 1984; Compton, Horn et al. 1989). Certainly, the modularity of rules allowed rules to be very easily added to such systems, but the addition of new rules invariably affected the prior performance of the system. Simply because a rule could be added in a modular fashion, gave no guarantee that it would not have other changes on the system apart from those intended. To deal with this problem, the research area of verification and validation emerged.

On approach to this problem was to control the way in which rules were added to minimise the impact on the system's performance (Compton and Jansen 1990). However, the major research focus in KBS development since the mid 80's, has been in attempting to understand these systems at the knowledge level and thereby develop more principled approaches to development. The knowledge level perspective points out that there is a whole level of inquiry about how domain knowledge is reasoned with to solve a problem a opposed to the content of the domain knowledge itself. If one goes back and considers earlier rule-based systems, one finds embedded in the knowledge base all sorts of different methods, or model-construction operators (Clancey 1992). The classic example of this, is the practice in MYCIN to put the conditions in a rule, in order of importance. There was no indication in the knowledge base for why this was done, but in fact, it was to ensure that the most significant data was dealt with first and that any questions asked of the user reflected this. The MYCIN backward chaining inference engine processed rule conditions in order of appearance in the rule. The aim of knowledge-level modelling has been to improve maintenance and re-use of PSMs by making such mechanisms explicit. The principle focus in this has been on tasks and then PSMs; i.e. what is the best way to reason with domain knowledge to find a solution to a particular type of problem. There have been many ways of designing, developing, organising and accessing PSMs, with the most influential of these from the mid-80's probably being the KADS methodology (Wielinga, Schreiber et al. 1992). For reviews and frameworks for understanding these modelling approaches see for example: (Brazier and Wijngaards 1997; Fensel, Motta et al. 1997). More recently ontologies have become a more major focus of knowledge-level modelling . If one considers ontologies in terms of Brachman's earlier work (Brachman and Schmolze 1985), it is fair to say that the present emphasis is on the T-Box, that is on definitions of terms and their relationships in an inheritance hierarchy. The other aspect of Brachman's perspective, the A-box, assertional knowledge, e.g. rules about the domain, is of little interest at present in knowledge acquisition (KA) research. From a knowledge-level modelling perspective, the content of the assertions in assertional knowledge is not part of the knowledge level and so is of little interest. However, the question in this paper is whether in practice there is a trade-off between PSM power and assertional knowledge about objects and properties in the domain. The assertional knowledge the expert is asked to provide is simply to identify the appropriate solution for a particular problem and identify the features of the situation which lead the expert to choses this particular solution, or component of the solution.

There has always been interest in flexible KBS where the reasoning of the system can adapted to the current problem e.g. (Brazier, Treur et al. 1995). However, more recently there has been an upsurge in interest in providing some more generic framework for understanding PSMs and their relationships. Fensel has developed a way of seeing PSMs as adaptations of each other (Fensel 1997). Perkuhn groups PSMs in terms of family resemblances (Perkuhn 1997). Significantly the core viewpoint in (Fensel, Motta et al. 1997) is that PSMs are search engines, but adapted to the particular task in hand. There is nothing novel in saying that PSMs are generalised search engines, however the historical evolution of this emphasis is of interest. AI started off with search methods. It was then realised that knowledge overcame many of the limitations of pure search. However, the use of knowledge was somewhat inchoate and a knowledge level perspective emerged to try to make sense of the use of knowledge. We are know back at a focus on search, but on understanding PSMs as different types of search, and the relationships between them related to the difference in search. The hypothesis in this paper is that there is advantage in continuing the cycle and returning to an emphasis on the notion of the utility of domain knowledge to reduce search. Although this knowledge reduces search, as will be seen, the domain expert is simply providing advice as to why a particular solution should be chosen in a particular context.

Finally in this perspective on PSMs, Beys sees PSMs as ways of mapping input sets to output sets (Beys and van Someren 1997). Again it is not novel to suggest that PSMs map inputs to outputs, however this re-emphasis at this stage is of interest. Beys uses a set based approach because of its mathematical utility. Our perspective is more simple in that essentially a KBS is a system that produces some form of output given some form of input and the output produced for any given input is the same as would be produced by an expert. The most primitive form of mapping between input and output is simply to connect inputs to the appropriate output, with every feature of the input being taken into account. This would be clearly inefficient as there would be no generality. If we consider a flat rule structure where a rule makes a final conclusion from the initial data, then rules are a mapping from input to output but where only the salient features in the case are identified. This provides greater generality but with this goes a greater liability for error when the generalisations are over-generalisations. If one then moves to something like heuristic classification, one attempts to provide some abstraction of the data and then rules relating to these abstractions. This is a further attempt at generalisation, or rather an attempt to minimise the amount to knowledge required and make use as much use of possible of knowledge added - essentially increasing the range in which given knowledge will be used. This attempt to increase the power of the knowledge by increasing the scope of where it might be used, again increases the possibility of errors (causing the maintenance problems we have already noted). Heuristic classification can be seen as a form of local search. Given some data there are other facts that can be concluded on the way to a solution. However, the facts that can be concluded will depend on both the knowledge available in the rules and the inference mechanism; how the conflict resolution strategy works and so on. If one moves to a more complex example such as the Propose and Revise PSM (Marcus and McDermott 1989; Zdrahal and Motta 1994) we have a system where there is some form of local search (propose), but when this develops a solution or partial solution that is not allowed because of some constraint, there is a mechanism to modify the current solution (i.e. to jump to some other part of the search space) and start the local search again. This can again be seen as a way of trying to generalise, or rather minimise the amount of knowledge required. The system is not provided with sufficient knowledge so that a local search will necessarily provide a solution. Rather when the local search fails the system is designed to jump to another part of the search space, to see if a solution can be found there. That is, it is trying to make the best use of the knowledge available.


We suggest that the way in which knowledge is categorised relates to the type of PSM that is proposed or anticipated, rather than any intrinsic requirement of the task. If we consider this in terms of configuration, or parametric design: Following (Motta and Zdrahal 1996), a configuration problem can be characterised as a mapping from a six-dimensional space <P, Vr, C, R, Pr, cf> to a set of solution configurations, {D1, D2, ..., Dn} where

P = parameters (attributes) = {p1, p2, ..., pn};
Vr = value ranges = {V1, V2, ..., Vn}, where Vi = { vi1, ..., vin};
C = constrains = {c1, c2, ..., cm};
R = requirements = {r1, ..., rk};
Pr = preferences = {pr1, ..., prj};
cf = global cost function.


Denoted by Dk = {<pi, vij>}, where pi [element] P and vij [element] Vi a configuration. This results in the following definitions:


We believe that this categorisation of knowledge follows from a Propose and Revise type of perspective. Certainly it also applies to other PSMs (Motta and Zdrahal 1996), but we would see these attempting the same types of advantage from generalising knowledge (or rather attempting to minimise the domain knowledge required) as Propose and Revise. As far as we can see there is no need to invoke a notion of constraint knowledge unless, one is planning on responding differently to a constraint violation than to an unsatisfied requirement. Similarly there is no need to understand requirements and preferences differently unless wishes to reason about them differently. The reason for requiring constraint type knowledge is to be able to test if constraints are satisfied and judge whether the search needs to be started in a new part of the space. If one considers problems simply in terms of mappings from inputs to outputs, then in parametric design tasks, the expert's core competence is simply in indicating what components are required in the solution for the given input and what features in the input indicate that these components are required. We also implicitly include in this, the notion that the expert will also provide information on when a solution is inappropriate for the given input and why a different solution is appropriate. A particular solution may be required or preferred because it satisfies constraints or is preferred over another solution, but in the situation the expert's judgement is simply that the features in the input indicate a particular solution is appropriate. Of course the expert may reflect on what he has done is all sorts of ways and provide all sorts of explanation, but his or her core competence is identifying features in the input which determine the output.

This then raises the question of a trade-off. Clearly one should attempt to make as much use of the knowledge available as possible. For example in Propose and Revise, if a solution is not found by some form of local search, the search is moved to another part of the space. On the other hand one can attempt to overcome the problem by adding further knowledge so that the problem is found by local search. In other words, for a parametric design problem, can a problem solver be developed by adding knowledge so that local search will always provide a solution, without a revise step? A similar approach was used in (Gaines 1992) where an assignment problem was solved by classification, but there was not the same emphasis on simplifying the search required by use of domain knowledge. Clearly it is likely that such a solution will probably require a greater amount of domain knowledge. On the other hand, generalisation produces errors and maintenance problems and in particular, despite its ability to move to new parts of the search space, Propose and Revise may not be very satisfactory at finding solutions (Zdrahal and Motta 1996). Zdrahal and Motta's solution (ibid) to this problem was to locate the search in an appropriate part of the space by using case-based reasoning. This is essentially an attempt to provide a solution by adding more domain knowledge. It is also an attempt to avoid the maintenance problems of dealing with the domain knowledge in the system, by providing another source of domain knowledge used in a different way.

AIM

The aim of this study is explore whether the addition of further domain knowledge can be used to extend the competence of a problem solver. This would also provide a significant contribution to the reuse possibility of PSMs as it would suggest that simpler PSMs could be used for more complex tasks. We have another underlying aim of extending Ripple Down Rules (RDR) to construction tasks, and in the first instance configuration tasks. RDR are a way of building knowledge based systems which focus on very simple addition of domain knowledge. The major success of RDR has been in single classification tasks. A challenge to RDR from the PSM community has been to extend RDR to a wider range of tasks than classification. As will be discussed below RDR have been extended to multiple classification tasks. They have also been used by others for control tasks (Shiraz and Sammut 1997) and for heuristic search (Beydoun and Hoffman 1997). This paper describes the extension of RDR to a configuration task. The advantage of RDR for exploring the trade-off hypothesis is that they facilitate knowledge acquisition and a trade-off between domain knowledge and PSM power can only be useful if domain knowledge can be added as required to enable the problems solver to find a solution.

Any RDR development depends on there being a sequence of cases which can be used to develop a system, so that RDR studies have generally been closely tied to specific applications. There is also little interest in developing RDR systems where there is not a genuine expert amiable. With other methods the knowledge engineering needs to study and understand the domain regardless of whether or not there is an expert. A consequence of this is that a knowledge engineer will make significant progress to at least developing a problems solver and framework for the domain knowledge without an expert to provide the knowledge. In contrast with RDR, after some fairly simple data modelling, there is no other task except for the expert to identify the features in the data for a case which indicate the appropriate solution. For this reason we did not attempt the Sisyphus II project. The application here is to develop an adviser for Ion Chromatography (IC) methods. The major advantages of this domain for this project is that we had access to a large database of cases as well as one of the world's experts on IC, and as well a there is real need and considerable interest from analytical chemists and equipment suppliers in developing such a system. IC is widely used as a standard tool to separate and thereby identify charged components of a solution, for example in pollution studies. The sample is injected onto a column and washed through with a solute. Different ions in the sample are retarded in their passage through the column. Factors affecting separation include: charge on the solute, polarizability of the solute and ion-exchange capacity of the stationary phase, hydrophobicity of the polymer used in the stationary phase, hydrophobicity of the solute, amount of organic material present, relative sizes of the solute and the pores of the stationary phase etc. Clearly, ion chromatography methods will vary with the composition of the sample being investigated and there are a number of different components of an IC method that have to be selected. The configuration task is to select an appropriate set of components and conditions to be able to measure the substances of interest in the sample at hand. This is a difficult task, and the last author has assembled a data base of some 4000 methods as a resource to assist in method selection for difficult problems. The particular sub-domain covered in the system developed so is Ion Exclusion Chromatography. Ion Exclusion Chromatography is concerned with separating acids and bases; however, exactly the same considerations, and the same range of options in configuring a method apply. In the cases covered to date by the system, separation of up to 16 different components is required.

The following contains a more detailed discussion of RDR, the Ion Chromatography domain, the RDR Configuration method and results of using this method on the IC domain.

RDR BACKGROUND

Ripple Down Rules (RDR) is based on the idea that when a KBS makes an incorrect conclusion the new rule that is added to correct that conclusion should only be used in the same context in which the mistake was made (Compton and Jansen 1990). In practice this means attaching the rule at the end of the sequence of rules that were evaluated leading to the wrong conclusion. Thus, this rule will only be evaluated in the same context in which the mistake was made. Rules are thus added when errors are made, allowing for an incremental, maintenance-focussed approach to knowledge acquisition. Since rules are added to correct the interpretation of a specific case, the system ends up not only with a number of rules but with cases associated with these rules: the cases that prompted the addition of the rule. This leads to validated knowledge acquisition in that the new rule that is added should not be allowed to be satisfied by any of the cases for which rules have already been added and for which different conclusions are appropriate. These cases are already handled.

For single classification problems, where the system will output one of a set of conclusions a binary tree structure is used and the new rule added after the last rule evaluated. It is worth noting that the tree is extremely unbalanced and resembles a decision list with decision list refinements rather than a tree (Catlett 1992). With this structure the only case that can reach the new conclusion is the case associated with the rule that gave the incorrect conclusion. Implicit in this outline is a non-monotonic approach in that each rule proposes a final conclusion to be output, unless this conclusion is replaced by a later correction rule that fires. Rules are never removed or edited, all errors are corrected by adding new rules.

A single classification RDR approach was used to build the PEIRS system for pathology interpretation. This system was largely developed whilst in routine use. Apart from the development of the initial domain model the knowledge base was developed entirely by an expert without knowledge engineering support or skills. The only task the expert had was to identify features in the case which distinguished it from the case associated with the rule giving the wrong conclusion and which was now being corrected. This allowed the expert to add about 2400 rules at the rate of about 3 mins per rule and the system ended up being in routine use for four years and covering about 25% of chemical pathology (Edwards, Compton et al. 1993). This compares very favourably with other medical expert system projects and with expert system projects in general in terms of knowledge acquisition tasks. A range of studies have also been carried out demonstrating that problems with repetition that may occur with this approach happen less in practice than might be expected (Compton, Preston et al. 1995; Compton, Preston et al. 1994).

The notion of repetition in the KB is an important one with respect to the trade-off between knowledge and PSM power. The use of a binary tree structure means that knowledge in one particular pathway is not accessible to a search down another pathway. Repetition in single classification RDR occurs because when a solution is not found down a particular path, further knowledge is added to that path rather moving the search to elsewhere. However. as noted. in practice the repetition is small, while the speed of knowledge acquisition with a 'chunk' added every three minutes meets the requirements of the new DARPA project on very large knowledge based systems. The requirement for this project is for KBS with between 10,000 and 100,000 chunks of knowledge to be built at this rate. PEIRS was only 2,500 rules, but the cost of knowledge addition was close to constant as the KB grew, where the DARPA requirement is close to linear with KB size, suggesting the DARPA requirement could be met.

For multiple classification tasks an n-ary tree structure was developed and replaced the binary tree structure used for single classification tasks. With this structure a rule may have many children and all children will be evaluated. If any of these are satisfied all their children will be evaluated and so on. In contrast for single classification, once a child rule was found to be satisfied by the data, only its children were evaluated, and as soon as of these were found to be true only the children of that rule would be evaluated and so on. In MCRDR all immediate children are evaluated regardless. We again have a simple non-monotonic system, where child conclusions replace the parent conclusion. As a result of the n-ary tree structure, cases associated with a number of other rules may reach a new rule being added. If a new child rule is being added, the cases associated with its parents, siblings and all the children of its siblings can all reach this rule, and so it must be made specific enough to exclude any of these cases satisfying this rule. It is assumed that all conclusions required are already provided by the KBS for these cases.

The approach to validate knowledge acquisition was simply to show the case to the expert and ask them to select features leading to the new conclusion. If these features were present in any of the cases that could reach the new rule, this case was also shown to the expert and the expert was required to add further features to the rule to exclude this case. This process was repeated until none of the 'cornerstone cases', cases associated with rules, were satisfied by the rule. The cornerstone cases associated with rules include the case for which the rule was added and also cases which correctly reach this rule but were stored because they caused the addition of another rule. Surprisingly even when up to 600 cases might reach the rule, the expert has to see only two or three cases before a sufficiently precise rule is constructed to exclude all cases. In various evaluation studies MCRDR has been shown to be more efficient that single classification RDR even for single classification tasks, but at the small cost of the expert having to deal with 2 to 3 cases for each rule added (Kang 1996; Kang, Compton et al. 1995).

ION CHROMATOGRAPHY

In ion chromatography, the set of parameters which specify a method are:

P = {ALCOHOL, ALDEHYDE, CHARGE, COL_PACK, COLUMN, CONDUCTANCE, DETECT_LIMIT, DETECTOR, ELECTRO, ELECTROACTIVE, ELECTRODE, FORMCOM, H_STATE, MOBILEPH_CONC, MOBILEPH_PH, MOBILEPHASE, MODIFIER, PKA, POST_COLUMN, SPECT, SUPPRESSED, TEMP, WAVELEN}

The range of values for these attributes are:

V(ALCOHOL) = {YES, NO}
V(ALDEHYDE) = {YES, NO}
V(CHARGE) = {NEG, PNEG, POS, PPOS}
V(COL_PACK) = {CATION_EX, LATEX}
V(COLUMN) = {HPICE_AS1, HPICE_AS2, HPICE_AS3, HPICE_AS4, HPICE_AS5}
V(CONDUCTANCE) = {YES, NO, SUPPRESSED}
V(DETECT_LIMIT) = {VERY_LOW, LOW, VERY_HIGH, HIGH, NORMAL}
V(DETECTOR) ={CONDUCT,AMPEROMETRY,PULSED_AMPER, FLUOROCENCE, UV_ABSOR, RI, POSTCOLM, DIR_SPEC}
V(ELECTRO) = {PULSED, AMPEROMETRY, NO}
V(ELECTROACTIVE) = {PU, AM, NO}
V(ELECTRODE) = { Pt, CARBON, Au, NON}
V(FORMCOM) = {MANNITOL, NON, PAR, MOLBD}
V(H_STATE) = {MONO, DI, TRI}
V(MOBILEPH_CONC) = {DILUTE, CONC, 1mM, 0.1M}
V(MOBILEPH_PH) = {ACIDIC, BASIC, NEUTRAL}
V(MOBILEPHASE) = {HCl, H3SO4, HNO3, H3O, MANNITOL, PFBA, OCT_SULFONIC, PERCHLORATE, P_FLOURHEPTAN, T_D_FLU_HEPT}
V(MODIFIER) = {ISOPROPANOL, NO}
V(PKA) = -10:100
V(POST_COLUMN) = {YES, NO}
V(SPECT) = {NON, FLUORCENCE, UV_ABSOR}
V(SUPPRESSED) = {YES, NO}
V(TEMP) = {ROOM_TEMP, 34}
V(WAVELEN) = {215, 200}


These parameters and values suggest about 1012 possible configurations. No doubt, the space of real possibilities is far smaller than this, but clearly the search space is quite sizeable.


Some example requirements for the domain:

C1 = IF (ANY(FORMCOM) = MANNITOL) THEN (MOBILEPHASE = MANNITOL)
C2 = IF (DETECTOR = UV_ABSOR) THEN (ELECTRODE=NO)
C3 = IF (MOBILE_PHASE = OCT_SULFONIC) THEN (DETECTOR = CONDUCT)


As noted, in the approach adopted we do not indicate separately: constraints, requirements and preferences. As will be seen, the expert simply indicates the features in the case which indicate a particular configuration is appropriate. Such rules can be considered as requirements. However, importantly, the approach also requires the expert to add rules to correct inappropriate configurations. Now these rules will still be requirement rules. However, since they cause the replacement of certain components of the previously proposed configuration, they also implicitly express constraints and preferences. There is also no global cost function as similarly costs are handled by the rules added.

A sample case for which the system may have to produce a method may look like:

SAMPLE


PHOSPHORIC


ARSENIC(V)


BORIC


PKA


2.13


2.19


9.24


CHARGE


PNEG


PNEG


PNEG


MW


98


0


61.84


RI


0


0


0


SPECT


NO


NO


NO


ELECTROACIVE


.


NO


NO


CONDUCTANCE


.


YES


YES


FORMCOM


MOLYBD



NO


H_STATE


TRI



MONO



The sample may contain only a single substance of interest, but this is a trivial case and of little interest. Here the sample contains three analyte ions and the method should be able to separate them all. In another example , the acid/base components of interest in white wine would be: Lactic, Tartaric, Malic, Acetic. In the cases considered to date separation of up to 16 components has been required. The system also contains a data base from which the information in the lower panels of the table is derived, e.g molecular weight, conductance etc. The development of this data base required a significant effort and it strictly speaking it is an important part of the problem solving. Initial attempts to develop a system, where components were identified by name only, were not very successful, because experts, even if only implicitly, reason about particular attributes of these components, and different combinations of components result in different combinations of values for these attributes. However, this database will not be considered further here as the major issue is incremental addition of heuristic rules given an appropriate data representation. Use of the data base corresponds to the abstraction stage in heuristic classification. Here however, the abstraction involved decomposition of what are essentially objects identified by a kind of proper name into their significant and more generic components. It should be noted that RDR do not claim to solve the problem of developing an appropriate data model for the domain. This remains a standard knowledge engineering, or rather software engineering task. However RDR do make this stage easier, because of the ability to refine over-simplified abstractions with reference to the raw data in local contexts (Edwards, Compton et al. 1993). Current projects are aimed at developing a more general RDR based facility for refining abstractions concurrently with knowledge that uses these abstractions.

The cases used in this study come from a data base of all the significant Ion Chromatography methods published between 1980 and 1996, amounting to over 4000 cases. Of these a smaller subset of Ion Exclusion Chromatography cases were selected These in fact comprised the first cases in the data base. An initial study developed configurations for a group of 81 cases which were referred to the last author for assessment. The assessment was simply on whether these methods would work for the particular sample. The configurations for a total of 361 cases were assessed by the first author as being equivalent to the specific methods in the data base. This assessment was necessary because often there are number of alternatives that will satisfy the same constraints and the selection of a particular one is unimportant, probably determined by what happened to be available in the laboratory. The second author is a PhD student in Chemistry, but not an expert in Ion Chromatography. His specific expertise in IC was developed during this project. He acted as the expert in providing rules for the expert system and as well in deciding that methods were adequate. Note that these two activities are integrated for RDR, in that rules are only added when the expert believes the solution presented is inadequate. The methods suggested for these cases will eventually be assessed by a more experienced expert. However, for this study the level of expertise is not of major importance. Note that the expert task here was not in deciding on a method, but in the more generic 'expert chemist' task of deciding whether a method proposed corresponded to a method in the data base for the identical case. Secondly the 'expert' had to gradually assemble rules to produce the specified method. It has previously been shown that RDR can be successfully used with 'lesser experts'. A high level of expertise can be achieved by the RDR system, but takes longer to develop as more correction rules have to added to correct earlier poor rules (Compton, Preston et al. 1995; Shiraz and Sammut 1997). If then a rapid development of expertise can be achieved with a PhD student expert, it is highly plausible that even better performance will be achieved with a more experienced expert.

CONFIGURATION RDR

Machine learning based system

This study follows an earlier project were machine learning was used to develop an Ion Chromatography expert system (Mulholland, Preston et al. 1996). The 4000 cases were used to develop a series of classification knowledge bases each able to decide on one of the components of the method given the other components and the other data for the case. These knowledge bases were build by induction with Gaines' Induct algorithm (Gaines 1991) being the method of preference because of its superior ability to deal with classes with small numbers of representatives.

Once these knowledge bases were developed the task was to develop a PSM that could use this knowledge to find a suitable solution. The method arrived at:

  1. applied the data to each of the KBS and added to working memory any conclusions, that is parts of the configuration that could be deduced from the data provided. The KBS were single classification so that only single conclusions for each attribute could be reached.
  2. Many of the rules in the various knowledge bases were unable to fire because the rules had been learnt from cases where the rest of the configuration was complete but were now being provided with very partial data. It was therefore assumed that all rules were satisfied, unless the data available explicitly excluded a rule from being satisfied. E.g. If there was rule (IF x = 42 THEN . . . .), this rule would be assumed to be satisfied if X was unknown but false if X = 40, for example.
  3. Step 2 allowed the possibility of multiple values for a particular configuration attribute. A heuristic was used that where a KBS provided a single choice for one of the configuration attributes, this added to working memory. Otherwise nothing was added to working memory for this attribute. The first two steps were then repeated. If no KBS could provide a single conclusion, the user was asked to select a conclusion to be added to working memory.
  4. The process stopped when a stable configuration was reached or a cycle in the sequence configurations produced was detected. It should be noted that at any stage of the inference any of the conclusions previously reached could be replaced, except for those arrived at on the first pass of step 1. That is, those parts of the solutions that could be arrived at directly from the external data without any assumptions about missing data or any dependence on conclusions derived during the reasoning process.
  5. The solutions was checked by seeing if each of the components of the solution could be derived by single classification reasoning from all the others. That is using the various KBS for the purpose they had been designed.


This problem solving method can be seen as a variant of propose and revise in that the heuristic of only adding components of the solution to working memory when only a single value for that component was possible, is a type of best first search. That this will be changed if the same process arrives at another conclusion corresponds to the revise part of propose and revise, This essentially starts the search in a new place. Since the requirement to move from one single conclusion to another is somewhat difficult to satisfy, convergence is reasonably good and the system does not jump around too much.

Manually developed knowledge bases

The machine learning based configuration system did not allow for incremental knowledge acquisition. To develop this we first of all moved to MCRDR to allow multiple conclusions. This did not offer any intrinsic advantage over multiple single classification knowledge bases. However, the n-ary tree structure with no false branches more transparently supported the possible requirement to hypothesise conclusions.

Initial efforts were focussed on adding knowledge acquisition to the type of PSM used for the ML configuration system but now applied to MCRDR. This proved to be very complex and involved tracing inference sequences and adding rules in such a way that the same inference sequence would be maintained for any case. The inference sequences tended to be quite long. While this was under investigation we attempted two simpler heuristics corresponding approximately to a best first and a complete search.
Heuristic 1: (ROSE)

Example:
process a case {a,b}
If we get {a, b, c1, c2, d}, i.e conflict in the values of parameter c
Remove conflict, we get {a, b, d}
Process the new case: {a, b, d}


Heuristic 2: (SALMON)

Example:
If we got {a, b, c1, c2}, i.e conflict in the values of parameter c
Generate all possible combinations
{a, b, c1}, {a, b, c2}
All cases are then processed recursively.
If processing a case results in a conclusion which conflicts with the case, then reject the case. E.g if {a, b1} lead to {a, b2} then reject {a, b1}.


We soon abandoned the SALMON strategy as too expensive and because ROSE was surprisingly effective.

The final inference approach used is as follows.

  1. Carry out a single pass MCRDR inference on a case (described in more detail below).
  2. If a single value is proposed for any attribute which does not yet have a value assigned, then add that attribute-value pair to working memory. Do this for all attributes for which a single value is proposed. Ignore attributes for which more than one value is found. Note that no attribute value pair assigned to memory is later retracted.
  3. Infer again and go back to 2.
  4. This cycle is repeated until either a complete configuration is reached or until no more single values can be found for remaining attributes. In this case the user is asked to select a single value from a single attribute from those inferred and the process returns to step 2.
  5. We anticipated that there would be situations where some attributes had no values inferred and that we would need to hypothesise that unknown conditions in the rule antecedent were in fact true as in the ML method above. We also anticipated that we would need some sort of revision step to move to other parts of the search space. However, these have not been necessary because the problems are overcome by further knowledge acquisition.


The knowledge acquisition strategy used is:

  1. If any component of the solution is wrong, or some component is missing, new rules are added (perhaps only one) to give the correct conclusion. This rule is added in the conventional MCRDR approach described in more detail below. The key feature of this is that inference during rule addition consists of a single pass. That is the only data that is already available is used by the system. As will be seen below, more than one rule may be added to ensure the single correct valued is assigned to an attribute.
  2. After the new rule is added to the knowledge base the case is re-run. If any further components are wrong, step 1 is repeated until finally a correct solution is delivered.


Propose and revise methods typically contain a forward chaining rule interpreter to cover the assignment of values to attributes, but then a process to revise this and move to a different part of the state space if the local search has reached an unsatisfactory solution. The central hypothesis here is that when such the system reaches an unsatisfactory or incomplete solution, it is better to add rules enabling it to find a solution rather than moving to a different part of the search space. Here the revision does not take place by moving to another part of the search space. Rather the domain knowledge is revised at acquisition time to improve local search.

In this method new knowledge is added to the system so that the single appropriate value for an attribute will be assigned. It is important to note that this assignment cannot be replaced by another value due to the simple addition of other rules to deal with other cases. Another rule firing will result in multiple values for the attribute and so that nothing will be added to memory for this attribute. For an alternate value to be assigned, a specific refinement rule, either stopping or replacing the earlier assignment has to be added, so that there will be only a single value for the attribute from the new rule. We believe this is a critical feature in the success of the approach. Also the knowledge acquisition ensures that as rules are added they will not be satisfied by the cases for which other rules have been added.

Multiple Classification Ripple Down Rules

Since configuration RDR depend critically on Multiple Classification Ripple Down rules, they are described here in some detail.

Inference

The MCRDR is represented as an n-ary tree. The inference of MCRDR evaluates all the rules in the first level of the KB, then evaluates the rules at the next level of refinement for each rule that was satisfied at the top level and so on. The process stops when there are no more children to evaluate or when none of these rules can be satisfied by the current data case. If any child rule fires, its conclusion replaces the conclusion of the parent rule. If more than one child fires, the conclusions of these rules are also included as conclusions, but perhaps to be replaced by conclusions from their children.

The inference process can be viewed as multiple paths from the root of the KB to the conclusions, as shown below in Fig 2.


Fig 1. An MCRDR knowledge base


Path 1 [(Rule 0, ...), (Rule 2, Class 2), (Rule 6, Class 6)]   Info 1
Path 2 [(Rule 0, ...), (Rule 2, Class 2), (Rule 10, Class 5)]  Info 2 [4]
Path 3 [(Rule 0, ...), (Rule 3, Class 2)]                      Info 3
Path 4 [(Rule 0, ...), (Rule 5, Class 5)]                      Info 4 [2]


Fig 2. Pathways through the KB from Fig 1. The rules producing conclusions are highlighted. "Info n [...]" indicates other rule numbers with the same classification.

Knowledge Acquisition

When a case has been misclassified or a new classification is required, knowledge acquisition is needed. The process has three aspects: firstly, the system acquires the correct classifications from the expert; secondly, the system decides on the new rules' location and finally, the system acquires new rules from the expert and adds them to correct the knowledge base.

Acquiring new classifications

The expert simply states the new classifications. For example, if the system produces classifications class 2, class 5, class 6 for a given case, the expert may decide that class 6 does not need to be changed but class 2 and class 5 should be deleted and class 7 and class 9 need to be added.

Locating rules

The system should find the locations for the new rules that will provide these classifications. There are several possibilities that are shown in the table below

Wrong classifications


To correct the KB


Wrong classification stopped


Add a rule (stopping rule) at the end of path to prevent the classification (gives a null classification)


Wrong classification replaced by a new conclusion


Add a rule at the end of path to give the new X classification.


A new independent classification


Add a rule at a higher level to give the new classification. In practice MCRDR systems add rules for independent classifications only at the top level.


Acquiring Rule Conditions - Rule Validation

With MCRDR, a number of cases can reach a new rule and the higher in the tree the more cases can reach the rule. The new rule should distinguish between new case and all of these cornerstone cases. The cases that can reach a rule are those associated with the parent of a new rule, its siblings and their children. Note that the cases associated with rule include cases for which any rule is added, but for which this rule gives a conclusion.

The algorithm for adding a new rule is as follows:

  1. The system assembles a list of all the stored case which can reach the new rule and should be distinguished from the current case.
  2. The expert is shown the case and one of these cornerstone cases and asked to select features which distinguish these cases.
  3. Cornerstone cases which do not satisfy this rule are removed from the cornerstone case list.
  4. The process is repeated from step 2 until there is no cornerstone case which satisfies the rule.


The same process is followed whether a stopping or refinement rule or new top-level rule is added.

RESULTS

Some sample rules follow. In the first a component of the configuration depends only on the data supplied. In practise the PKA of the various ions in the sample is derived from the data base and then rules can refer to these abstractions of the data.

IF (MIN(PKA) >= 2.0) & (MAX(PKA) <= 4.0) THEN MOBILE_PHASE = HCl

The following is a rule which has conditions derived from other components of the configuration.

IF (DETECTOR= CONDUCT)&(MOBILE_PHASE = HCl) THEN SUPPRESSED= YES


Fig 3. The IC system provided configurations for 361 cases. The continuous line shows the number of cases for which the configuration was not acceptable. Rules were added to fix all erroneous configurations. This histograms show the number of rules added per case.

DISCUSSION

The key feature of the results is that the rate at which errors are made on cases, and therefore the rate at which rules have to be added, rapidly decreases. Although the system has not been evaluated on a large independent set of cases, the cases on which the system is tested here are random cases which have not been preselected. They were in fact the first 361 Ion Exclusion Chromatography cases chronologically in the data base. It is therefore reasonable to assume that error rate on these cases should be an approximation of the error rate more generally. Since errors are being corrected as they occur, the average over any sequence of cases should be greater than the actual error at the end of the sequence with the errors corrected. The error rate on the first 100 cases was ~17%, it was ~7% for the second 100 cases and <3% for the last 100 cases. This final error is after only 361 cases have been seen and 156 rules have been added. The convergence for what might be assumed to be a difficult configuration problem is quite remarkable. From experience with PEIRS and other RDR systems, the system will never reach completion and rules will continue to be added for cases throughout the 4000 in the data base. This is also consistent with experience with machine learning, that knowledge in a system is never complete (Catlett 1991).

That there are further errors to be corrected may seem to be a weakness of the method. However, evaluation of expert systems always indicates some level of error and the error here after only 361 cases is comparable to good final error rates of other expert systems. The difference here however is that the possibility of further errors is fully recognised and accepted and provision is made for fixing these as they occur. Generally in conventional systems, the published error rate is the final performance level of the system and there is little facility for correcting errors except careful editing of the knowledge base followed by separate verification and validation.

The second significant aspect of this development is the number of rules that have to be added for each case corrected. In the initial stages rules have to be added for virtually every component of a configuration. However as the system develops, cases which require a lot of rules become fewer and fewer, again showing the improvement of the system. Of interest, occasionally the later cases require many rules, suggesting that these case are quite different from those already handled. One would expect this phenomenon to continue, but with decreasing frequency.

We have already argued that because of the type of expertise required in developing this system that it was reasonable to use a Chemistry PhD as the expert. A real expert will only produce a better system. We in fact provided 81 cases to the fourth author and developer of the IC data base. Of these 88% of the methods were identified as workable and only 12% as not workable. The main area of error was in choosing a mobile phase and showed up a weakness in the second author's knowledge about the interchangeability of acids in IC.

This evaluation is not complete and it is therefore possible to raise questions about the quality of the system as it develops further. It may also be hypothesised that the search space in reality is far smaller than it seems. All we can say is that this hypothesis conflicts with the experience of analytical chemists in setting up Ion Chromatography methods for particular samples. This is seen as difficult task where the average practitioner may need expert advice or a series of false starts before an appropriate method is set up. The next stage in this project is to distribute both a knowledge base and tools for further knowledge addition to about six laboratories outside Australia recognised internationally for their leadership in Ion Chromatography. It seems unlikely that these laboratories would be keen to participate if this was a small problem. Despite these indications to the contrary we remain open to the suggestion that we have in fact provided a solution to a small problem. However, it is also fair to say, that compared to the level of evaluation available for other problem solving methods for configuration tasks, configuration RDR is exceptionally well validated. One may well ask for more validation of configuration RDR, but it would seem unreasonable to argue that one should prefer other less well validated methods because of the possibility that the validation of configuration RDR may not be complete.

One significant conclusion from this study is that we have now shown that the RDR incremental development approach can be extended to construction tasks. This opens up the possibility of incremental knowledge acquisition for the whole range of knowledge based system tasks. The next major issue in this research is to integrate RDR and refining data models for a domain, rather than data modelling being a conventional knowledge engineering exercise. A first step in this is integrating a subsumption hierarchy and RDR so that changes to neither the A-box or the T-box should degrade the performance of the system (Martinez, Benjamin et al. 1997). As we have noted there is other independent work on using RDR for control tasks and heuristic search. There have also been a number of independent uses of RDR as representation in machine learning.

We believe the central issue raised by these results is the trade-off between knowledge acquisition and problem solving. As Motta and Zdrahal noted there is considerable variety in the various implementations of Propose and Revise (Motta and Zdrahal 1996). This suggests that researchers were attempting to come up with or tailor optimal PSMs for such problems. Most PSM literature is written as if the domain knowledge is somehow available and the issue is how to develop a PSM which can use this domain knowledge to find a solution. We suggest that this leads to the idea, at least implicitly, that if a PSM cannot find the solution, then a better PSM is needed. In contrast RDR suggests that more knowledge should be added which the PSM can use to find the solution. In RDR, when knowledge, relevant to the case in hand is not found, rather than fix the PSM, more knowledge is added that can be found with the given PSM.

The key question is whether this confers any advantage. On the face of it the advantage is considerable. A configuration RDR PSM can be used for multiple classification. There will simply be no inference cycles because conclusions reached do not affect other conclusions reached. We have also shown that single classification is better carried out with an MCRDR PSM. This means that a single PSM can be used for these three tasks. No doubt a Propose and Revise PSM could also be used for classification but there are major differences in how knowledge is categorised in these methods. In heuristic classification there is an abstraction step and then a refinement step. In Propose and Revise there is both a local search step and a revision step moving to another part of the space. There is also a clear distinction between the knowledge used in the different stages. In contrast with RDR all knowledge is simply an identification of salient features in a case which lead to a particular part of the output required. There is no need to understand the internal structure. However, the internal structure is virtually identical for MCRDR and configuration RDR. The long standing of claim of RDR is that this type of incremental validated knowledge acquisition is rapid and reliable and something that can be carried out by an expert without the aid of knowledge engineer. In this regard it is worth noting that neither the expert nor anyone else in the project knows anything about the actual organisation of the knowledge in the IC knowledge base.

The question for RDR of course remains whether RDR PSMs can be developed or the current PSM extended to deal with still further KBS tasks. We ourselves would see the extension to a configuration task as a major step in demonstrating that they can be used for construction tasks. Secondly, we see the present results as suggesting that perhaps these apparently more difficult problems can be approached in much simpler ways if greater use is made of domain knowledge. However, our own experience in developing such methods is that progress is slow. Firstly RDR PSMs require the major extra feature that they support incremental validated knowledge acquisition. However, the problems in coming up with such PSMs were that we invariably attempted to develop something that was far more complex than is actually required.

ACKNOWLEDGMENTS

This work has been supported by an Australian Research Council Grant. The insightful and helpful comments of the anonymous referees are greatly appreciated although we doubt we have fully addressed the concerns they raised

REFERENCES

Bachant, J. and McDermott, J. (1984). R1 revisited: four years in the trenches. The AI Magazine Fall: 21-32.

Beydoun, G. and Hoffman, A. (1997). Acquisition of search knowledge, in E. Plaza and R. Benjamins (Eds.), Knowledge Acquisition, Modelling and Management. Berlin, Springer. 1-16.

Beys, P. and van Someren, M. (1997). A systematic approach to the functionality of problem solving, in E. Plaza and R. Benjamins (Eds.), Knowledge Acquisition, Modelling and Management. Berlin, Springer. 17-32.

Brachman, R. and Schmolze, J. (1985). An overview of the KL-ONE knowledge representation system. Cognitive Science 9(2): 171-216.

Brazier, F. and Wijngaards, N. J. E. (1997). An instrument for a purpose driven comparison of modelling frameworks, in E. Plaza and R. Benjamins (Eds.), Knowledge Acquisition, Modelling and Management. Berlin, Springer. 323-328.

Brazier, F. M. T., Treur, J., Wijngaards, N. J. E. and Willems, M. (1995). Formal Specification of hierarchically (de)composed tasks. Proceedings of the 9th Banff Knowledge Acquisition for Knowledge-Based Systems Workshop, Banff, SRDG Publications, Dept. of Computer Science, University of Calgary. 25.1-25.20

Brownston, L., Farrell, R., Kant, E. and Martin, N. (1985). Programming Expert Systems in OPS5. Reading, Mass, Addison-Wesley.

Catlett, J. (1991). Megainduction: A Test Flight. Machine Learning: Proceedings of the Eighth International Conference., Morgan Kauffman. 596-599

Catlett, J. (1992). Ripple-Down-Rules as a Mediating Representation in Interactive Induction. Proceedings of the Second Japanese Knowledge Acquisition for Knowledge-Based Systems Workshop, Kobe, Japan, 155-170

Clancey, W. J. (1992). Model construction operators. Artificial Intelligence 53(1-115):

Compton, P., Horn, R., Quinlan, R. and Lazarus, L. (1989). Maintaining an expert system, in J. R. Quinlan (Eds.), Applications of Expert Systems. London, Addison Wesley. 366-385.

Compton, P., Preston, P. and Kang, B. (1995). The Use of Simulated Experts in Evaluating Knowledge Acquisition. Proceedings of the 9th AAAI-Sponsored Banff Knowledge Acquisition for Knowledge-Based Systems Workshop, Banff, Canada, University of Calgary. 12.1-12.18

Compton, P., Preston, P., Kang, B. and Yip, T. (1994). Local patching produces compact knowledge bases, in L. Steels, G. Schreiber and W. Van de Velde (Eds.), A Future for Knowledge Acquisition: Proceedings of EKAW'94. Berlin, Springer Verlag. 104-117.

Compton, P. J. and Jansen, R. (1990). A philosophical basis for knowledge acquisition. Knowledge Acquisition 2: 241-257. (Proceedings of the 3rd European Knowledge Acquisition for Knowledge-Based Systems Workshop, Paris 1989, pp 75-89)

Edwards, G., Compton, P., Malor, R., Srinivasan, A. and Lazarus, L. (1993). PEIRS: a pathologist maintained expert system for the interpretation of chemical pathology reports. Pathology 25: 27-34.

Fensel, D. (1997). The tower-of-adaptor method for developing and reusing problem-solving methods, in E. Plaza and R. Benjamins (Eds.), Knowledge Acquisition, Modelling and Management. Berlin, Springer. 97-112.

Fensel, D., Motta, E., Decker, S. and Zdrahal, Z. (1997). Using ontologies for defining tasks, problem-solving methods and their mappings, in E. Plaza and R. Benjamins (Eds.), Knowledge Acquisition, Modelling and Management. Berlin, Springer. 113-128.

Gaines, B. (1991). Induction and visualisation of rules with exceptions. 6th AAAI Knowledge Acquisition for Knowledge Based Systems Workshop, Banff, 7.1-7.17

Gaines, B. (1992). The Sisyphus problem-solving example through a visual language with KL-ONE-like knowledge representation. Sisyphus'91: Models of Problem Solving, GMD (Gesellschaft fur Mathematik und Datenverarbeitung mbH).

Kang (1996). Multiple Classification Ripple Down Rules. UNSW, PhD thesis.

Kang, B., Compton, P. and Preston, P. (1995). Multiple Classification Ripple Down Rules : Evaluation and Possibilities. Proceedings of the 9th AAAI-Sponsored Banff Knowledge Acquisition for Knowledge-Based Systems Workshop, Banff, Canada, University of Calgary. 17.1-17.20

Marcus, S. and McDermott, J. (1989). SALT: A knowledge Acquisition Language for Propose-and-Revise Systems. AI Journal 39: 1-38.

Martinez, R., Benjamin, R., Preston, P. and Compton, P. (1997). Integrating Ontology Development and Knowledge Acquisition.

Motta, E. and Zdrahal, Z. (1996). Parametric Design Problem Solving. Proceedings of the 10th AAAI-Sponsored Banff Knowledge Acquisition for Knowledge-Based Systems Workshop, Banff, SRDG Publications, University of Calgary.

Mulholland, M., Preston, P., Haddad, P., Hibbert, B. and Compton, P. (1996). Teaching a Computer Ion Chromatography from a Database of Published Methods. Journal of Chromatography 739: 15-24.

Perkuhn, R. (1997). Reuse of problem-solving methods and family resemblances, in E. Plaza and R. Benjamins (Eds.), Knowledge Acquisition, Modelling and Management. Berlin, Springer. 174-189.

Shiraz, G. and Sammut, C. (1997). Combining Knowledge Acquisition and Machine Learning to Control Dynamic Systems. 15th International Joint Conferences on Artificial Intelligence (IJCAI'97) 1997., Nagoya, Japan,

Wielinga, B. J., Schreiber, A. T. and Breuker, J. A. (1992). KADS: a modelling approach to knowledge engineering. Knowledge Acquisition 4(1 (special issue on KADS)): 5-54.

Zdrahal, Z. and Motta, E. (1994). An In-Depth Analysis of Propose and Revise Problem Solving Methods. Proceedings of the Third Japanese Knowledge Acquisition for Knowledge-Based Systems Workshop (JKAW'94), Tokyo, Japanese Society for Artificial Intelligence. 89-106

Zdrahal, Z. and Motta, E. (1996). Improving Competence by Integrating Case-Based Reasoning and Heuristic Search. Proceedings of the 10th AAAI-Sponsored Banff Knowledge Acquisition for Knowledge-Based Systems Workshop, Banff, SRDG Publications, University of Calgary.