THE INFERENCE EFFICIENCY PROBLEM IN BUSINESS
AND TECHNOLOGICAL RULES MANAGEMENT SYSTEMS
Barbara Baster and Andrzej Macioł
Faculty of Management, AGH University of Science and Technology, ul. Gramatyka 10, 30-067 Kraków, Poland
Keywords: Business Rules Management Systems, Technology, Rule-based Systems, Inference Efficiency, Databases.
Abstract: In the following paper we present the results of our works related to development of rule engine for
automated interpretation of business rules. Our experiences and experimental results show that knowledge
description for this purpose may be stored in the form of relational databases. The aim of the experimental
research presented in this paper was to determine degree in which organization of knowledge base and
assumed inference strategy influence the efficiency of inference process itself. Experiments proved that
owing to the application of mechanism characteristic for relational data bases, knowledge base can be easily
arranged so as to maximize efficiency of inference. The efficiency of inference is strongly influenced by
preliminary knowledge transformation from the set of examples or random rules into arranged form.
1 INTRODUCTION
In the paper we present the results of our ongoing
works related to development of rule engine for
automated interpretation of business rules. Our
researches are connected with the project cofounded
by the European Union. The works are aimed at
development of a tool which will operate in
accordance with general BRM idea and will
simultaneously include specific technological
knowledge.
Recent researches highlight the particular
specificity of issues of the borderline of management
and technology. Because of this specificity, the
business rule management systems that are currently
available on the market are very rarely applied in the
fields related to process and technology
management. One of the factors that contributes to
the current state of matter is a necessity to apply
hybrid system in this field – connection of rule-
based system with procedural computations. Such
situation takes place e.g. in case of the technology
development for a new product, or process
management in emergency situations. In such
instances, a rule-based system would control the
selection process of conditions that should be
verified depending on the results of preceding
computations. The verification of individual
conditions would be based on execution of
computations in accordance with predefined
algorithms (procedural computations). Similarly,
hybrid systems could support the processes related
to the disposal of faulty products.
During research works it was also found that
typical decision problems connected with
technology issues do not require preparation of
complex knowledge models. Nevertheless, the
number of rules may in some cases grow to
hundreds and complicate the acquisition of
knowledge simultaneously raising the requirements
above the limits settled by inference systems.
In this publication we present the results of
research which was aimed at assessment of
possibilities to improve the inference procedure
from the point of view of processed queries and time
consumption. Moreover, one of our goals is to prove
that concept which utilizes relational database as
vocabulary carrier gives wide range of possibilities
in the field of improvements of knowledge structure
and the efficiency of inference process.
2 INFERENCE EFFICIENCY
PROBLEM
At present, many forward-chaining systems use
inference networks to capably match rule conditions
to facts. The most efficient rules engines are
designed by the Rete algorithm. It is a widely-used
match algorithm for comparing a large numbers of
10
Baster B. and Macioł A. (2010).
THE INFERENCE EFFICIENCY PROBLEM IN BUSINESS AND TECHNOLOGICAL RULES MANAGEMENT SYSTEMS.
In Proceedings of the 5th International Conference on Software and Data Technologies, pages 10-15
DOI: 10.5220/0002931600100015
Copyright
c
 SciTePress

patterns to a large collection of slowly changed data
without iterating over the sets (Forgy, 1982). The
algorithm was originally developed by Charles
Forgy (1981) for OPS5 and it has soon become the
basis for many popular system shells, including for
example NASA’s CLIPS (Giarratano, 1989), Jess or
Drools.
The Rete algorithm is designed to increase the
speed of forward-chaining rule systems. Systems
based on Rete construct a network of nodes. Each
node corresponds to a pattern occurring in the
condition part of a rule (the left-hand-side) and
contains a memory of facts which fulfil that pattern.
When facts are changed, they propagate along the
network and cause nodes to be annotated if the fact
matches the pattern. The output of Rete is a set of
new matched patterns, and a set of formerly matched
patterns no longer matching (Lee and Schor, 1990).
Research involving Rete algorithm has been
focused on improving its performance, as in (Kang
and Cheng, 2004), (Zhou et al., 2008), (Tuttle and
Eick, 1991). Tuttle and Eick (1991) propose a
generalization of the Rete network, called a
historical Rete network which can store and
maintain a run’s historical information.
In the literature except for Rete we can also find
different production rule condition testing
algorithms. Wang and Hanson (1993) present one of
them. Their paper shows the results of a simulation
study comparing the join-condition testing efficiency
of Rete and TREAT (a variation of Rete) in the
context of a database production rule system.
Actually TREAT is very similar to Rete. The way
that both algorithms test selection conditions is
identical. The difference lies in how they test joins.
According to the results of Wang and Hanson’s
research TREAT algorithm in many ways
outperforms Rete. It requires less storage than Rete,
and is less sensitive to optimization decisions. Based
on these results, authors conclude that TREAT is the
preferred algorithm for testing join conditions of
database rules, however, they agree that both of
these two match algorithms can be successively used
to improve the speed of the match process in
DBRSs.
3 APPLICATION
OF RELATIONAL DATA BASES
AS A KNOWLEDGE MODELS
FOR BRE
Our experiences indicate that the most effective
solution, capable of direct cooperation with majority
of industrial information systems which
simultaneously provides decidability, is a
combination of relational model with inference
system that utilizes attributive logics. Our solution,
named Inference with Queries (IwQ) (Macio,
2008), has been developed as a knowledge model
and an inference engine for formulation of Business
Rules Management Systems. Knowledge storage is
realized in accordance with principles of Variable
Set Attribute Logic (VSAL).
An assumption was made that in rule-based
decisive system, attributes as well as variables will
be stored in form of relations. In our solution, we
have introduced data metamodel. Concept of data
metamodel comes down to partial transfer of
intensional database structure to its extensional part.
Solution based on a relational database and
metamodel allows for very simple and flexible
development of both data structure and inference
process. Owing to that fact final effects resemble
those obtained by means of Rete algorithm, yet some
specific redundancy characteristic for this solution is
excluded.
4 TECHNOLOGICAL
PROBLEMS EXAMPLES
Our research works concentrated on the field of
metal processing in its broad sense, but their final
results may be generalized. Despite vast diversity
and uniqueness of business processes that are
realized in the analysed field, there exist multiple
processes that can be relatively easily automated.
Those processes are conducted in a cyclical manner
and realized on the basis of relatively simply
definable quantitative and qualitative criterions. In
the case of non-measurable criterions a common
practise is to asses point-like scores. These processes
can be described with the use of rule-based systems.
The processes which are repeatable and their
frequency of realization is relatively high should be
automated in the first place.
In spite of the fact that similar problems are
solved by means of the BRM systems which are
widely available on the market, in our case we face
the problem of connecting data that originates from
different sources. In dependence of facts’ source, we
may predict different level of costs, hence the
inference strategy becomes an important matter.
Institutions that verify contractor’s credibility
constitute an exemplary case of such ‘costly’ source
of information. It becomes obvious that rules which
THE INFERENCE EFFICIENCY PROBLEM IN BUSINESS AND TECHNOLOGICAL RULES MANAGEMENT
SYSTEMS
11

require such type of data should be avoided at all
cost as long as remaining rules allow for final
settlement (e.g. in case when quality of material
offered by the contractor – supplier is not
satisfactory, there is no use in verification of his
credibility).
5 EXPERIMENT
5.1 Aim and Assumptions
The aim of the experimental research was to
determine the degree in which organization of
knowledge base and assumed inference strategy
influence the efficiency of inference process itself.
In the first step an analysis was conducted
whether structure of knowledge base, understood as
number of rules and their connection pattern,
influences inference time. BRM systems most often
do not provide mechanism of knowledge generation
on the basis of examples (which arises from the
assumption, that knowledge in rules management is
a priori defined). In our opinion, in considerations
where decision tables are utilized we deal with a set
of examples, that can be transformed by means of
machine learning methods into set of rules in
simpler form than the set corresponding to every
column of the table. Owing to this transformation
the number of rules can be reduced, which entails
streamlining of the inference process.
In our research works we analysed to what
extent does the transformation of exemplary sets
described below to the form of decision table will
allow to increase efficiency of inference.
Transformations were performed by means of
selected algorithm of decision tree learning (ID3).
Method of knowledge description in the form of
database, which was elaborated by our team, allows
for extremely simple manipulations on accessibility
conditions to the rules and facts, without a necessity
of changes in the inference algorithm. Such changes
may be introduced by altering the parameters of
these SQL queries which pass information to the
inference engine. In our research we were looking
for the answer to the question whether modifications
in access sequence for the rules and facts does
increase efficiency of inference in case of such
structure of the knowledge base. For this purpose we
conducted experiments both for unsystematic
knowledge as well as different variants of its
systematisation. We did analyse what are the
consequences of systematization within the tables
storing rules and facts.
In a vast majority of currently available BRM
systems a forward chaining inference is used. The
inference engine receives either a single piece of
information or a set of information items about facts
and begins with classic forward chaining. In case
when the data is not complete and firing of the rule
is impossible, the engine queries for missing
premises.
Our system’s character resembles inference
system in a mixed manner, it works in accordance
with classical scheme which combines forward
chaining and backward chaining for the search of
missing premises. When the general definition of
inference target is prepared, a start point must be
defined, i.e. attribute which values must be inputted
explicitly at the very beginning of the process.
We have made an attempt to define rules that
would allow to determine which of the source
attributes (that are not the result of inference) should
be selected as the starting point. For that purpose we
conducted simulations which enabled us to settle the
dependencies between selection strategy of starting
point and the inference efficiency.
Simulation analysis of inference systems
requires preliminary settling of the rules for
answering these questions, which constitute a basis
for verification of facts correctness. In case of a
system that is founded on attribute logic, this process
is based on selection of particular feature’s value of
the attribute from the domain of this feature. In the
basic research we made an assumption, that ever
value of the feature is equally probable. This
assumption corresponds to a situation when a
knowledge engineer does not have an opportunity to
asses variability of attribute values. These
assumptions were taken for initial works.
Because of the fact that feature values of
attributes are not uniformly distributed, we have also
analysed impact of this phenomena on efficiency of
inference process.
5.2 Knowledge Base
The inference schema is shown in Figure 1. It is
based on relational database which contains the
assessment criteria as it shows Figure 2. The three-
degree and two-degree qualitative scale has been
used to estimate individual premise (good, average,
bad or yes and no).
Sample rule is as follows:
Rule: 5
if
Certainty = bad
Quality = yes
Logistics = good
ICSOFT 2010 - 5th International Conference on Software and Data Technologies
12

then Supplier := average
5.3 Research Plan
Procedure of the experiments was aimed at
observation of inference process in above-described
database for randomly generated responses. Number
of queries of the inference engine as well as process
duration time were the objects of observation. Since
the number of rules was significantly limited,
additional retarding element was introduced which
prolonged answering time of queries (without it,
firing times of about 1 millisecond were observed in
extreme cases).
Figure 1: An inference schema.
Supplier - Conclusions
Economy Logistics Quality Certainty SUPPLIER
good good no good average
good good no average average
good good no bad average
good good yes good good
good good yes average good
good good yes bad average
good average no good average
good average no average average
Figure 2: Criteria of supplier selection (partial).
Considering the aim of the research, the
following schedule of experiments was prepared:
1. Inference in a raw knowledge base, in which
the number of rules corresponded to the
number of examples (E1)
2. Inference in a knowledge base after its
transformation into decision tree by means of
ID3 algorithm (the number of rules was
reduced from 90 to 60) (E2)
3. Inference in the transformed knowledge base
and after rearrangement of the rules according
to increasing number of premises, for each
separate attribute that constituted conclusion,
for 3 randomly selected variants of fact
arrangement inside the database (E3, E4, E5)
All of the above-mentioned experiments were
conducted under assumption of equal probability of
each value that an attribute can take.
Moreover, one additional experiment was
performed for the most favourable variant among all
previously described, but under assumption of
dispersed probability distribution of attribute values
(E6).
5.4 Experimental Results
Table 1: Comparison of results for the experiments E1
and E2.
Attribute Number of queries
E1 E2 BP*
TRANSPORT 8 7,30 8,75%
EXTENSIVENESS 8 7,16 10,50%
CERTAINTY 8 7,14 10,75%
STRENGTH 8 7,14 10,75%
OTHER FEATURES 8 7,12 11,00%
PRICE 8 7,18 10,25%
PAYMENT CONDITIONS 8 7,36 8,00%
RECURRENCE 8 7,66 4,25%
MEAN 8 7,26 9,28%
Process duration
E1 E2 BP*
TRANSPORT 621,25 580,00 7,11%
EXTENSIVENESS 618,59 563,13 9,85%
CERTAINTY 684,69 621,25 10,21%
STRENGTH 569,06 506,25 12,41%
OTHER FEATURES 572,97 506,56 13,11%
PRICE 624,22 566,56 10,18%
PAYMENT CONDITIONS 619,38 577,81 7,19%
RECURRENCE 624,53 593,13 5,30%
MEAN 684,69 621,25 9,42%
* betterment in percent
the best betterment result in bold
Transforming the set of examples into non-
arranged decision tree results in reduction of number
of queries by 9.28% on average and reduces the
inference consumption time by 9.42%. Obviously,
the number of asked queries in the raw knowledge
base is not dependent on the start point. Even after
its transformation into decision tree, no influence of
the start point selection on the number of queries
asked can be noted. On the other hand, selection of
THE INFERENCE EFFICIENCY PROBLEM IN BUSINESS AND TECHNOLOGICAL RULES MANAGEMENT
SYSTEMS
13

the starting point does influence inference time in
the raw as well as transformed knowledge base.
In the table 2. an juxtaposition is included, where
results of experiments for the raw knowledge base
are compared with the mean result obtained after the
arrangement of rules for three different arrangement
patterns of fact table.
Arrangement of the rules has a very distinct
influence on efficiency of inference process.
Independently on the selection of start point, a mean
time of inference is reduced by 22,38%, and the
number of asked queries falls by 19,32%. Also, the
most favourable start point is clearly distinguishable.
It is the fact which directly co-decides about final
conclusion and simultaneously reveals itself to be
certainty fact. If the inference is started from this
very point, the inference time may be reduced by
23,55% with respect to the best possible time
obtained for the raw knowledge base. At the same
time, number of asked queries decreases by 30,25%
with regard to the maximum possible number.
Table 2: Comparison of results for the experiments E1
and mean result of experiments E2, E4, E5.
Attribute Number of queries
E1 E3,E4,E5 BP*
TRANSPORT 8 6,22 22,25%
EXTENSIVENESS 8 6,30 21,25%
CERTAINTY 8 5,58 30,25%
STRENGTH 8 6,97 12,88%
OTHER FEATURES 8 6,88 13,96%
PRICE 8 6,12 23,50%
PAYMENT
CONDITIONS
8
6,09 23,88%
RECURRENCE 8 7,47 6,62%
MEAN 8 6,45 19,32%
Process duration
E1 E3,E4,E5 BP*
TRANSPORT 621,25 469,06 24,50%
EXTENSIVENESS 618,59 476,56 23,29%
CERTAINTY 684,69 460,57 25,86%
STRENGTH 569,06 493,54 20,56%
OTHER FEATURES 572,97 479,48 22,82%
PRICE 624,22 457,76 26,32%
PAYMENT
CONDITIONS 619,38 438,02 29,49%
RECURRENCE 624,53 582,60 6,22%
MEAN 684,69 482,20 22,38%
* betterment in percent
the best result in bold
In the table 3. an juxtaposition is included, which
depicts comparison of results for the experiments
conducted on the knowledge base with arranged
rules for three different arrangement patterns of the
fact table.
The results of experiments imply that the
arrangement patter of facts does not have significant
influence on mean inference time, number of asked
queries or the most profitable selection of start fact.
Table 3: Comparison of results for the experiments E3, E4
and E5.
Attribute
Number of queries
E3 E4 E5
TRANSPORT 6,40 6,30 5,96
EXTENSIVENESS 6,40 6,18 6,32
CERTAINTY 5,42 5,62 5,70
STRENGTH 6,91 7,08 6,92
OTHER FEATURES 6,91 6,90 6,84
PRICE 6,08 5,94 6,34
PAYMENT CONDITIONS 6,17 6,10 6,00
RECURRENCE 7,61 7,40 7,40
MEAN 6,49 6,44 6,44
Attribute Process duration
E3 E4 E5
TRANSPORT 489,69 480,00 437,50
EXTENSIVENESS 496,56 459,06 474,06
CERTAINTY 447,97 461,88 471,88
STRENGTH 503,13 493,75 483,75
OTHER FEATURES 489,69 477,50 471,25
PRICE 462,03 435,94 475,31
PAYMENT CONDITIONS 449,06 435,31 429,69
RECURRENCE 594,06 571,88 581,88
MEAN 491,52 476,91 478,16
* betterment in percent
the best result in bold
Table 4. contains comparison of average values
for different experiments. One for arranged
database, conducted under assumption of uniform
distribution probability of attribute values and the
remaining for the same database, but with dispersed
probability.
As it can be clearly seen, the variable character
of attribute values does not influence the inference
time or the number of asked queries. It may be
concluded from the fact, that optimization of
knowledge base structure can be executed during its
design phase. Thus there is no necessity to analyze
the character of described phenomena in a statistical
point of view.
6 CONCLUSIONS
Experiments conducted during our research have
proven that owing to application of mechanism
ICSOFT 2010 - 5th International Conference on Software and Data Technologies
14

characteristic for relational data bases, knowledge
base can be easily arranged so as to maximize the
efficiency of inference. It also became clear that the
efficiency of inference is strongly influenced by
preliminary knowledge transformation from the set
of examples or random rules into arranged form (e.g.
form of decision tree). Arrangement of the set of
rules as well as selection of the start point are also
significant issues. By proper arrangement of the
rules and selection of the favourable start point,
inference time can be reduced by 23,55% and the
number of asked queries decreased by 30,25%.
Table 4: Comparison of the mean result of experiments
E2, E4, E5 and experiment E6.
Attribute Number of queries
E3,E4,
E5
E6 BP*
TRANSPORT 6,22 6,22 0,00%
EXTENSIVENESS 6,30 5,96 5,40%
CERTAINTY 5,58 5,86 -5,02%
STRENGTH 6,97 7,00 -0,43%
OTHER FEATURES 6,88 6,74 2,03%
PRICE 6,12 6,06 0,98%
PAYMENT CONDITIONS 6,09 6,72 -10,34%
RECURRENCE 7,47 7,04 5,76%
MEAN 6,45 6,45 0,00%
Attribute Process duration
E3,E4,E
5
E6 BP*
TRANSPORT 469,06 464,06 1,07%
EXTENSIVENESS 476,56 446,25 6,36%
CERTAINTY 460,57 496,25 -7,75%
STRENGTH 493,54 488,75 0,97%
OTHER FEATURES 479,48 470,63 1,85%
PRICE 457,76 457,81 -0,01%
PAYMENT CONDITIONS 438,02 492,81 -12,51%
RECURRENCE 582,60 546,88 6,13%
MEAN 482,20 482,93 -0,15%
* betterment in percent
the best result in bold
The results presented in this article prove that the
possibility to improve the inference efficiency
without necessity of pattern matching algorithms
application exists. Even though, they should be
treated as an introduction to further research on this
matter. In the experiments described above, the
simulation was applied as a tool for finding of quasi-
optimal solutions. Nonetheless, the simulation itself
can not be conducted in every specific case.
Therefore, the aim of our further research will be to
create the rules formulation that will allow to
achieve the best parameters for the structure of
knowledge base.
ACKNOWLEDGEMENTS
This research has been partially supported by the
Innovative Economy Operational Programme EU-
founded project (UDA-POIG.01.03.01-12-163/08-
01).
REFERENCES
Forgy, C.L. (1982). Rete: A Fast Algorithm for the Many
Pattern/Many Object Pattern Match Problem. Artificial
Intelligence. 19, 17-37.
Forgy, C.L. (1981). OPS5 User's manual. Technical
Report. Department of Computer Science, Carnegie-
Mellon University.
Giarratano, J. (1989). CLIPS User’s Guide, Version 4.3,
Artificial Intelligence Section, Lyndon B. Johnson
Space Center, COSMIC, 382 E. Broad St., Athens.
Kang, J.A., Cheng, A.M.K. (2004). Shortening Matching
Time in OPS5 Production Systems. IEEE
Transactions on Software Engineering, 30.
Lee, H.S., Schor, M. (1990). Dynamic Augmentation of
Generalized Rete Networks for Demand-Driven
Matching and Rule Updating. The 6th Conf. of A.I.
Applications 1990 IEEE.
Lee, Y.H., Yoo,S.I: (1995). A Rete-based Integration of
Forward and Backward Chaining Inferences.
Monterey.
Macioł, A. (2008). An application of rule-based tool in
attributive logic for business. Expert Systems with
Applications. 34, 1825–1836.
Tuttle, S.M., Eick, C.F. (1991). Historical Rete Networks
for Debugging Rule-Based Systems. San Jose, CA
(1991).
Wang, Y.W., Hanson, E.N. (1993). A Performance
Comparison of the Rete and TREAT Algorithms for
Testing Database Rule Conditions. Washington.
Zhou, D., Fu., Y., Zhong, S., Zhao, R. (2008). The Rete
Algorithm Improvement and Implementation. 2008
International Conference on Information
Management, Innovation Management and Industrial
Engineering.
THE INFERENCE EFFICIENCY PROBLEM IN BUSINESS AND TECHNOLOGICAL RULES MANAGEMENT
SYSTEMS
15