Applied Intelligence 21, 301–310, 2004

c 2004 Kluwer Academic Publishers. Manufactured in The United States.

A Fuzzy Integral Based Query Dispatching Model in Collaborative

Case-Based Reasoning

SIMON C.K. SHIU, YAN LI AND FENG ZHANG

Department of Computing, The Hong Kong Polytechnic University, China
csckshiu@comp.polyu.edu.hk
csyli@comp.polyu.edu.hk
cszhangf@comp.polyu.edu.hk

Abstract. In a collaborative (distributed) Case-Based Reasoning (CBR) environment, an input query case could
be compared with the old cases that are resided in many different CBR agents in the network. How to obtain
the best solution effectively and efficiently from this distributed CBR network depends on a carefully designed
query dispatching strategy. In this paper, we propose a fuzzy integral based approach to measure the competence
of different CBR agents in the network and suggest three query dispatching policies which could be used to fulfill
this task. They are: To-Top policy, Strong-Strong policy and Best-Committee policy. The experimental result shows
that our proposed policies are comparatively better than the existing ones developed by Plaza and Ontañón.

Keywords: fuzzy integral, query dispatching, collaborative case-based reasoning

1. Introduction Plaza and Arcos [2] proposed two modes of cooper-

Traditionally, intelligent systems are developed in a
standalone and insolated manner. However, the recent
growth of the World Wide Web and multi-agent sys-
tems triggers the need of designing intelligent systems
in a distributed and collaborative manner. Being a suc-
cessful intelligent system technology, Case-Based Rea-
soning (CBR) also has the need to develop its applica-
tions into a full fledged and distributed environment.
Currently, there are two main approaches for select-
ing CBR agents, the first one is proposed by Nagendra
Prasad et al. [1] based on the concept of task decom-
position, and the second one is proposed by Plaza and
Arcos [2] based on the random selection of agents. The
use of task decomposition is only effective when the
problem can be nicely decomposed into a set of sub-
problems, and each sub-problem can be solved by an
individual CBR agent. However if conflicts exist (e.g.
the solutions from two sub-problems could not be in-
tegrated), additional heuristics from the users may be
needed. Sometimes this way of collaboration may be
even worse than a single and isolated system [3].
ation among CBR agents (i.e. Distributed Case-based
Reasoning (DistCBR) and Collective Case-based Rea-
soning (ColCBR)). DistCBR means that a problem can
be dispatched to any agent for solving, disregarding
who generates the problem. ColCBR means that the
owner of the problem tries to collect the useful cases
and methods from other agents, and decide how to solve
the problem. Three collaboration policies (i.e. Commit-
tee policy, Peer-Counsel Policy and Bounded-Counsel
Policy) were developed for the DistCBR framework by
Plaza and Ontañón [4]. However, these policies are all
based on a random selection of the agents, therefore in
most cases, it is very time consuming to get the best
solution.
Instead of having random selection of agents, we
propose our policies based on the concept of “compe-
tence”, which could be defined as the range of prob-
lems that a particular agent (or case) could solve. Since
these policies are based on “competence”, they are
comparatively better than the previous ones. The struc-
ture of this paper is as follows: Section two reviews
two methods of calculating the cases competence. The
302 Shiu, Li and Zhang
first one is proposed by Smyth and Keane [5] while
the second one is proposed by the authors. The CBR
agents competence computation and ranking policy are
given in Section 3. In Section 4, three policies for
dispatching a new query case are proposed. Each of
these policies is based on a different assumption of
how to obtain the best solution. An experimental com-
parison of our approaches to the existing ones is pro-
vided in Section 5. Finally, the conclusion is given in
Section 6.
2. Case-Base Competence Modeling
The concept of case-based competence was first pro-
posed by Smyth and Keane [5], (i.e. refer as the S-K
model in this paper), and subsequently it has been de-
veloped further to a whole range of concepts which
are useful for measuring the problem solving abil-
ity of case-bases. In the S-K model, many statistical
properties of a case base, such as the size and den-
sity of cases, are used as input parameters for mea-
suring competence. However, this model assumes that
there is no overlap among different group of cases (e.g.
features interaction [6] is a common cause of over-
laps). Therefore, if simply taking the group compe-
tence as the sum of the individual case competence,
and each individual case competence is computed inde-
pendently without considering the overlapping effects,
the resulting group competence may be over- or under-
exaggerated. This feature overlap problem has been
tackled by Shiu et al. [7] using fuzzy integral (refer as
the S-L model in this paper). These two models will be
used as the basis to develop our query case dispatch-
ing strategies. Details are explained in the following
sections.
2.1. The S-K Model
In this model, two key fundamental concepts are used:
coverage and reachability. Coverage of a case refers to
the set of problems that the case can solve. Reachability
of a case is the set of cases that can be used to provide
solutions to a case. Furthermore, the competence of a
group of cases (G) (i.e. group coverage of G) depends
on the number of cases in the group and their density
(see Eq. (1)).
GroupCoverage(G)
= 1 + |G| · (1 − GroupDensity(G)) (1)
where GroupDensity is defined as the average
CaseDensity of the group (see Eq. (2)), and |G| means
the size of the competence group G, i.e. the number of
cases in the group G.


GroupDensity(G) = CaseDensity(c, G)/|G|
c∈G
(2)
where CaseDensity is defined by Eq. (3)
CaseDensity(c, G)


= Simlarity(c, c∗ )/(|G| − 1). (3)
c∗ ∈G−{c}
Different ways of computing Similarity can be used de-
pending on the problem on hand. For a given case-base,
with competence groups G = {G 1 , G 2 , . . . , G n }, n =
1, 2 . . . , the total coverage is defined by Eq. (4).


Coverage(G) = GroupCoverage(G i ) (4)
G i ∈G
2.2. Problem of the S-K Model
Suppose that in some problem domain, we have a group
of non-uniformly distributed cases as depicted in Fig. 1,
it can be shown that the S-K model is not a good pre-
dictor of the group competence because this model as-
sumes that the cases are distributed uniformly such
as those shown in Fig. 2. Assuming that Size(G) =
Size(G ), i.e. |G| = |G  |, and GroupDensity(G) =
GroupDensity(G  ). Then, from Eq. (1), we have
GroupCoverage(G) = GroupCoverage(G  ) where G
is an arbitrary competence group in a case-base, so sim-
ilar results can be obtained between two case-bases, in
which one has its cases non-uniformly distributed and
the other uniformly distributed.
However, from Figs. 1 and 2, it is obvious that the
coverage of the two competence groups cannot pos-
sibly be the same. There are coverage holes in Fig. 1
compared with that of Fig. 2. If we calculate the com-
petence of the groups in Fig. 1 using the S-K model,
then the actual competence will be over-exaggerated.
It is because, the S-K model only considers the group
density, but ignores their distribution. There are possi-
bly many ways of cases distributions, therefore a more
accurate way of modeling of case-base competence is
required.
To further illustrate this point, we use the following
example:
 A Fuzzy Integral Based Query Dispatching Model
Figures 1 and 2. (1) Non-uniformly distributed case base. (2) Uniformly distributed case base.
Suppose that the densities of the groups G 1 and
G 2 in Fig. 1 are both 0.8 and they are assumed to
have uniform distribution (i.e. the case density of each
case, in either G 1 or G 2 , is 0.8). The density of the
whole group is 0.2, and the coverage of c∗ is three
cases. The overlap coverage of c∗ and G 1 ∪ G 2 is
two, and c∗ is a pivotal case. It is rather straightfor-
ward to get the coverage of the whole competence
group G as follows: (note that GroupCoverage1 means
the computed group coverage in Fig. 1 while Group-
Coverage2 means the computed group coverage in
Fig. 2):
GroupCoverage1(G) = GroupCoverage(G 1 )
+ GroupCoverage(G 2 ) + [coverage(c∗ )
− coverage(c∗ ) ∩ Coverage(G 1 ∪ G 2 )]
= 1 + [|G1|(1 − GroupDensity(G))] + 1
+ [|G2|(1 − GroupDensity(G))] + 1
= 1 + 5(1 − 0.8) + 1 + 7(1 − 0.8) + 1
= 5.4
However, according to the S-K competence model
(Eq. (1)), the result becomes:
GroupCoverage2(G) = 1 + [|G|
· (1 − GroupDensity(G))]
= 1 + 12(1 − 0.2)
= 10.6
The above two results are very different and the prob-
lem is caused by the inaccurate assumption that group
G is uniformly distributed, which is not true. This
303
also resulted with a computing error that cannot be
ignored.
In cases distribution such as Fig. 1, the difference
between GroupCoverage1 and GroupCoverage2 can
be further investigated as follows:
GroupCoverage2 − GroupCoverage1
= {1 + [|G| · (1 − GroupDensity(G))]}
− {1 + [|G1| · (1 − GroupDensity(G1))] + 1
+ [|G2| · (1 − GroupDensity(G2))] + 1}
= |G| · [(GroupDensity(G1)
− GroupDensity(G)) − 1 − GroupDensity(G1)]
(5)
≥ |G| · [(GroupDensity(G1)
− GroupDensity(G)) − 2] (6)
Given the above Eqs. (5) and (6), if the number of
cases increases and tends to ∞ (in the extreme case),
the value [GroupDensity(G 1 )−GroupDensity(G)] also
increases at the same time. As a result, the computing
error also tends to ∞. Therefore, the precision error of
Smyth and Keane’s competence model can be arbitrar-
ily large in some case-bases.
2.3. Fuzzy Integral Competence Model
In both Figs. 1 and 3, we can easily see that the case
c∗ or c∗∗ has an important role to play because they
will affect the overall competence distribution in the
group. Therefore, it is important to detect such cases
(which are called weak-links in this paper) for possible
identification of smaller competence groups which are
having more evenly distributed cases. These smaller
304 Shiu, Li and Zhang
Figure 3. An example of non-uniform distribution of cases in a
case-base.
groups’ competence can then be computed using the
S-K model. A new way of computing the group com-
petence (referred as the S-L model in the paper) is then
developed to tackle the above problem. Details are as
follows.
In general, competence groups, such as G 1 and G 2 in
Fig. 1, are not necessarily having strictly uniform dis-
tribution, and the weak link case c∗ is not necessarily a
pivotal case. To deal with this situation, without influ-
encing the results in Eq. (5), GroupDensity(G 1 ) can be
replaced by the average group density of group G 1 and
G 2 , which can be denoted by GroupDensity(G i ) i ∈
{1, 2}, So [GroupDensity(G 1 ) − GroupDensity(G)]
is equal to [GroupDensity(G i ) − GroupDensity(G)]
which is denoted by
GroupDensity. A concept called
quasi-uniform distribution can be used to describe the
distribution which is near to uniform distribution. As
According to inequality (8), since Relative
Coverage(c∗ ) is small, we can see that it is
Group
Density which leads to the competence error.
2.3.1. Detecting Weak Links. In order to tackle the
problem of non-uniformly distributed cases, we should
first identify the weak links in each competence group.
The definition of weak link as well as several related
concepts are more directly related to the competence
of the group in question, and are defined as follows:
Definition Let G = {c1 , c2 , . . . , cn } be a given com-
petence group in a case-base C, c∗ ∈ G is called a weak
link if
CompetenceError(c∗ )
= |G|
GroupDensity − GroupDensity(G i )
− RelativeCoverage(c∗ )
≥ α
where α is a parameter which depends on the ques-
tion on hand. If ∃c∗ ∈ G, and c∗ is a weak link, then
the competence group G is called a non-uniform dis-
tributed competence group. Otherwise, if for ∀c ∈ G,
CompetenceError(M) ≤ α, then G is called a quasi-
uniform distributed competence
 group.
It is obvious that G i = G − {c∗ }. With this def-
inition, we propose a recursive method to detect the
weak links in a given competence group G, which is
described as follows:
Weak-link Detection Algorithm:
1. W -SET← { }, G-SET ← { }, i = |G|;
mentioned, the other assumption that c∗ is a pivotal case
in the example is not necessarily true in many cases. To
address this problem, just consider the individual com-
petence of c∗ as its relative coverage, which is defined
as follows (Eq. (7)):
RelativeCoverage(c)

K competence model, add c∗ to the set of weak-links
1
= (7)
c ∈CoverageSet(c)
|ReachabilitySet(c )|
Hence, according to Eqs. (5) and (6), that is
Competence-error(c∗ )
= |G|
GroupDensity − Gr oup Densit y(G i )
− RelativeCoverage(c∗ ) ≥ |G|
GroupDensity
− (RelativeCoverage(c∗ ) + 1) (8)
2. If (i = 0) {Consider each given competence
group G in the S-K competence model, compute
CompetenceError(c), ∀c ∈ G; i = i − 1}
3. If there is no weak link, add G to G-SET, end;
4. If there is a weak link c∗ , identify the competence
groups G 1 , . . . , G n , (n ≥ 1) in G −{c∗ }using the S-
W -SET.
5. For (1 ≤ i ≤ n){G ← G i ; repeat Steps 1 to 4}.
Thus, we can obtain the set of weak links W -SET in
a given competence group G and the set of new compe-
tence groups G-SET. Then, a given competence group
G is repartitioned (i.e. divided into smaller groups) by
identifying the weak links in it. The groups in G-SET
are called new competence groups.
 A Fuzzy Integral Based Query Dispatching Model
2.3.2. Computing the Overall Coverage Using Fuzzy
Integral. After detecting the weak links in a com-
petence group G, several new competence groups
G 1 , . . . , G n (n ≥ 1) are produced. According to the
definition of a weak link, each newly produced group
is said to be quasi-uniformly distributed. The next task
is to compute the overall coverage of G. In the example
described in Fig. 1, we simply sum the competence of
G i (1 ≤ i ≤ n) and the relative coverage of c∗ , but this
method is only suitable in simple situations. There are
more complicated scenarios, such as the one given in
Fig. 3. It is difficult to clearly identify the contribution
of each weak link, (i.e. it is difficult to tell whether c∗
has more influence on the coverage of G than c∗∗ or vice
versa.) To describe this complex relationship, we apply
a powerful tool called fuzzy integral (or non-linear in-
tegral) with respect to a fuzzy measure (a non-additive
set function). Details are described in the next section.
2.3.3. Non-Additive Set Function. Let X be a
nonempty set and (X ) be the power set of X . We
use the symbol µ to denote a non-negative set func-
tion defined on (X )with the properties µ() = 0. If
µ(X ) = 1, µ is said to be regular. When X is finite, µ is
usually called a fuzzy measure if it satisfies monotonic-
ity, i.e., A ⊆ B ⇒ µ(A) ≤ µ(B) for A, B ∈ (X ).
For a non-negative set function µ, there are some as-
sociate concepts. µ is said to be additive if µ(A ∪ B) =
µ(A) + µ(B) for A, B ∈ (X ); to be sub-additive
if µ(A ∪ B) ≤ µ(A) + µ(B) for A, B ∈ (X ); to
be super-additive if µ(A ∪ B) ≥ µ(A) + µ(B) for
A, B ∈ (X ).
Let X = {G 1 , . . . , G n } be the space of the new com-
petence groups, and A and B two subsets of the power
set of X . Here, A and B can be a single new group G i or
the joint of several groups. If we consider µ(A) as the
importance of subset A, then the additivity of the set
function means that the joint importance of the groups
is just the sum of their respective importance, which
implies that there is no interaction among the compe-
tence groups. However, this is not true in the problem
considered. In fact, most measures of importance are
non-additive.
The sub-additivity and super-additivity are two spe-
cial types of non-additivity. Super-additivity means that
the joint importance of the two sets is greater than or
equal to the sum of their respective importance, which
indicates that the two sets are enhancing each other. In
contrast, sub-additivity means that the joint importance
of two sets is less than or equal to the sum of their re-
305
spective importance, which indicates the two sets are
resisting each other.
In our problem, consider X = {G 1 , . . . , G n } as the
factor space. There are weak links among the compe-
tence groups, linking them to one group G. Here, weak
links such as c∗ and c∗∗ are enhancing the overall cov-
erage of G. Hence, the important measure µ defined
on the power set (X ) is a super-additive measure. So
here we have
µ(A ∪ B) ≥ µ(A) + µ(B) for A, B ∈ (X ).
For example, in Fig. 3, c∗ enhances the importance of
G 1 ∪ G 2 , c∗∗ enhances the contribution of G 2 ∪ G 3 , and
there is no case to enhance or reduce the contribution
of G 1 ∪ G 3 , so we have
µ(G 1 ∪ G 2 ) ≥ µ(G 1 ) + µ(G 2 )
µ(G 2 ∪ G 3 ) ≥ µ(G 2 ) + µ(G 3 )
µ(G 1 ∪ G 3 ) = µ(G 1 ) + µ(G 3 )
When using the fuzzy integral to compute the over-
all coverage of the original competence group G, we
should determine the importance measure µ in ad-
vance. However, for a factor space including n factors,
there are (2n − 1) parameters to decide. In the situa-
tion of Fig. 3, seven values of the importance measure
should be determined, which are:
µ(G 1 ), µ(G 2 ), µ(G 3 ), µ(G 1 ∪ G 2 ), µ(G 1 ∪ G 3 ),
µ(G 2 ∪ G 3 ), µ(G 1 ∪ G 2 ∪ G 3 ).
To reduce the load, we apply a kind of fuzzy measure
called the λ-fuzzy measure, which takes the following
form:
µ(A ∪ B) = µ(A) + µ(B) + λ · µ(A) · µ(B)
λ ∈ (−1, ∞)
If λ ≤ 0, µ is a sub-additive measure; if λ ≥ 0, µ is
a super-additive measure; if and only if λ = 0, µ is
additive. So here we have λ ≥ 0. Applying the λ-fuzzy
measure to determine the importance measure µ, we
simply need to determine the importance of n on each
single factor and λ.
306 Shiu, Li and Zhang
2.3.4. Determining the λ-fuzzy Measure µ. In our
model, we consider that the importance of each compe-
tence group is equal to 1, i.e. µ(G i ) = 1, (1 ≤ i ≤ n).
This is a reasonable assumption because each group
makes a unique contribution to the overall coverage,
that is, the status of each group is considered to be
equal.
The next task is to determine the parameter λ, which
is critical to determine µ. It is obvious that the prop-
erties of the weak links between two groups are im-
portant factors for determining λ. In our model, cover-
age of a group refers to the area of the target problem
space covered by the group. In this sense, the value
of λ is closely related to the coverage of weak links
and the density of their coverage sets. Consider two
arbitrary new groups G i and G j , the W-SET between
them is C ∗ = {c1∗ , . . . , ch∗ }. We define Coverage(C ∗ )
and Density(C ∗ ) as follows:


h
Coverage(C ∗ ) = RelativeCoverage(ci∗ )
i=1
 ( f 1 , f 2 , . . . , f n );


h
∗
Density(C ) = GroupDensity(Cov(ci∗ )) h (2) Computing
i=1
where Cov(ci∗ )
is the coverage set of one of the weak
links between G i and G j .
The coverage contribution of G i ∪ G j must be di-
rectly proportional to Coverage(C ∗ ) and inversely pro-
portional to Density(C ∗ ). With these assumptions, the
parameter λ is given by the formula in Eq. (8).
λ = Coverage(C ∗ ) · (1 − Density(C ∗ )) (9)
The λ-fuzzy measure µ is then determined.
2.3.5. Using the Choquet Integral to Compute Com-
petence. Due to the non-additivity of the set function
µ, some new types of integrals (known as non-linear
integrals) have to be used. A common type of nonlin-
ear integrals with respect to non-negative monotone set
functions is the Choquet integral.
Let f be a nonnegative real-value measurable func-
tion defined on X , and µ be a non-negative monotone
set function introduced in the above section. The
 Cho-
quet integral of f on X with respect to µ, (c) f dµ,
is defined by the formula
  ∞
(c) fdµ = µ(Fα ) dα,
0
where Fα = {x | f (x) ≥ α} for any α ∈ [0, ∞). When
X is finite, the Choquet integral can also be defined in
the same way with respect to a non-negative set func-
tion that is not necessarily monotone.
In our model, X = {G 1 , . . . , G n } is finite, f i =
GroupCoverage(G i ), and importance measure µ satis-
fies:
µ(G i ) = 1(1 ≤ i ≤ n);
µ(A ∪ B) = µ(A) + µ(B) + λ · µ(A) · µ(B)(λ ≥ 0),
where λ is determined by Eq. (9).
The process of calculating the value of the Choquet
integral is as follows:
(1) Rearranging { f 1 , f 2 , . . . , f n } into a non-decrea-
sing order such that
f 1∗ ≤ f 2∗ ≤ · · · ≤ f n∗
where ( f 1∗ , f 2∗ , . . . , f n∗ ) is a permutation of


n
(c) fdµ = [ f j∗ − f j−1
∗
] · µ({G ∗j , G ∗j+1 , . . . , G ∗n })
j=1
where f (x0∗ ) = 0.
The value of the Choquet integral is considered as the
coverage of the considered competence group. Each
competence group in the S-K model is considered in
the same way, and the sum of all group coverage is the
overall coverage of the given case-base.
3. Competence of the CBR Agents
In the previous section, we have given the fuzzy integral
approach for calculating the competence of each case
group, in this section, we can compute the competence
of the CBR agents using the S-K and the S-L models
respectively. The policies proposed here are based on
the DistCBR mode (see section one) in which different
CBR agents are able to communicate and cooperate
with one another for recommending a solution. For in-
stance, when agent Ai is unable to solve a problem,
will delegate its authority of solving the problem toA j .
For a given Collaborative Case-Based Reasoning
(CCBR) system, there are n(n ≥ 1) case-based reason-
ers, denoted by CBR1 , CBR2 , . . . , CBRn . These CBRs
 A Fuzzy Integral Based Query Dispatching Model 307

can be regarded as n agents A1 , A2 , . . . , An for prob- 2. Compute the competence of each CBR agent in
lem solving in a distributed manner. The corresponding the CCBR2 group according to the S-L model,
case-bases are CB1 , CB2 , . . . , CBn , with competence and rank them as C12 , C22 , . . . , Cm2 2 .
groups G 1 , G 2 , . . . , G n respectively.
Step 4. Rank the CBR agents according to the respective
competence of each CBR agent in the CCBR1 and
Compute the Group Competence
CCBR2:
In computing the competence, we define the similarity According to the competence, rank the CBR agents in
between two cases p and q by the following equation: the CCBR1 and CCBR2 system in a descending order
  

as {A11 , A12 , . . . , A1m 1 }, {A21 , A22 , . . . , A2m 2 }.
n
S M pq = 1/ 1 +
(x pj − xq j )2  ,
j=1
4. Query Dispatching Policies
where xi j corresponds to the value of feature F j (1 ≤
Three query dispatching policies based on the compe-
j ≤ n), (i = 1, . . . , n).
tence of the CBR agents are proposed here: To-Top
policy; Strong-Strong policy and Best Committee pol-
Step 1. Detecting the weak-links in the above compe-
icy. These are:
tence group G i (i = 1, 2, . . . , n):
If ∃c∗ ∈ G i , s.t. CompetenceError(c∗ ) ≥ α, then the
competence group G i is a non-uniform distributed 4.1. To-Top Policy
competence group. Otherwise, G i is a quasi-uniform
distributed competence group. The main idea of this policy is to choose the CBR
Step 2. Partition the CBR agents according to their com- agent which has the maximal competence in the cor-
petence: responding CCBR system, i.e. A11 or A21 in CCBR1 and
the CCBR2 system respectively. CCBR1 is chosen as
1. CCBR1 ← φ, CCBR2 ← φ, i = |G|; where
the problem-solving agent if there are no feature in-
CCBR1 consists of those agents who have no
teractions, otherwise CCBR2 is chosen. For example,
feature interactions (or no overlaps among com-
in a travel-planning problem which will be described
petence groups), while CCBR2 consists of those
in section five, the hotels are classified by the number
agents who have feature interactions (or over-
of stars, therefore when the user specify the type of
laps among competence groups).
accommodations (e.g. the number of stars), this will
2. If (i = 0), compute CompetenceError(c), ∀c ∈
G, i = i − 1;
3. If there is no weak-link in G, then G is called
a quasi-uniform distributed competence group,
then add G to CCBR1, otherwise G is called
a non-uniform distributed competence group,
then add G to CCBR2, end;
4. For (1 ≤ j ≤ n), G ← G j ( j = 1, 2, . . . , n),
repeat the above Steps 1 to 3.
Step 3. Compute the competence of each CBR agent in
the CCBR1 & CCBR2:
1. Compute the competence of each CBR agent in
the CCBR1 group according to the S-K model.
Since the cases are distributed uniformly in
each CBR agent, we can get the competence
using the S-K competence model [5] directly.
They are then ranked in a descending order ac-
cording their competence, and are denoted as
C11 , C21 , . . . , Cm1 1 .
limit the choices of the available hotels. In this case,
the features “accommodation” and “hotel” are inter-
acting. The dispatching procedure is as follow: if agent
Ai receives an input query, it will try to solve it. When
the solution is satisfactory (i.e. within a user defined
threshold of solution accuracy, and efficiency), it be-
comes the answer to the input query. Otherwise, it will
dispatch the problem to A11 or A21 for solving. If Ai
is one of the agents in CCBR1, then it dispatches the
problem to the agent A11 , otherwise the problem will go
to A21 in CCBR2.
4.2. Strong-Strong Policy
In this policy, we assume that we could not determine
whether the features are having interactions or not.
Consider the travel-planning problem again, we are not
sure whether there are feature interactions or not be-
tween the features “season” and “hotel” or between
308 Shiu, Li and Zhang
the features “holiday duration” and “season”. Thus, it
is better to ask more than one agent to suggest the so-
lutions. We will choose the most competent agent (i.e.
one from each collaborative CBR system). That is, if
j
groups (i.e. each group represent one CBR agent, there-
fore if 4 agents are used, each of them consists of 200
cases, etc.). The feature “price” is chosen as the solu-
tion feature. The testing is based on the evaluation of
the agent Ai (i = 1, j = 1, 2) receives the prob-
lem, and cannot solve it satisfactorily, it will ask the
agents A11 in the CCBR1 and agent A21 in the CCBR2
to solve it in parallel. One of these suggested solutions
will be used based on an earlier assessment of these
two agents’ ability. (Note that these two agents belong
to the two CCBR systems, therefore the selection has
already considered the feature interaction property).
4.3. Best-Committee Policy
If time is not critical, and getting better solution is the
main concern, then the user can ask more agents for
suggested solutions. In general, the more the agents
are involved, the more accurate the answer will be.
j
the solution accuracy and the mean cost of solving (i.e.
time consumption).
The objective of the experiment is to determine the
“price” of each travel plan using our proposed policies.
A comparison of our approach to some exiting ones
[4] is also carried out. The mean relative error (i.e.
the difference between the actual result and the pre-
dicted result and divided by the actual result) is used
to compute the accuracy. For example in our experi-
ment, if four agents are used to predict the “price” of a
particular testing case (such as Case number 987), our
three policies will give the following results respec-
tively: $4,708.25, $3,536.72, and $4,561.35. The ac-
curacies are 84.94%, 86.43% and 88.26% respectively
(see Fig. 4). The mean cost is the relative CPU time of
the isolated agent, and assuming the mean time cost of
That is, if the agent Ai (i = 1, j = 1, 2) receives
the problem, it could follow the To-Top and Strong-
Strong policies first for solving the problem. However,
if the solution is not satisfactory, it can ask the agents
j j j
A1 , A2 , . . . , Ai−1 , (i.e. those agents that are better in
competence), to solve the problem. Each agent will of-
fer a solution to the problem, and the final solution is
chosen according to the user’s preference, such as pre-
ferred accuracy. This policy provides the user a flexible
choice, when he wants to get the best solution, then he
can ask all the agents for suggestions. This policy is
the same as the “Committee” policy proposed by Plaza
and Ontañón.
the isolated agent is ONE unit, then the mean time costs
of the collaborative policies are given in Fig. 5. We have
conducted five testing runs, and each testing has dif-
ferent number of agents. These agents are formed by
randomly re-organize the 800 testing cases.
The result shows that our policies use less time and
still can achieve the same accuracy as the other exist-
ing ones. In details, Fig. 4 shows that all of six case
dispatching policies are better than the isolated agent.
The Strong-Strong policy is better than the To-Top pol-
icy and the Best-Committee policy is the best one.
The Strong-Strong policy has similar accuracy to the

5. Experimental Evaluation

In order to evaluate our approach, we used a set of

test cases that are available from AI-CBR web-site (i.e.
www.ai-cbr.org). This travel case-base contains 1470
cases, and we randomly selected 1100 cases for our ex-
periment. Each test case describes a holiday-package
tour from Europe/ North Africa, and consists of 9 fea-
tures. In dividing the cases into different groups for
measuring competence, we use a random selection ap-
proach. The reason is that because in real life, there
may be missing values in the cases, therefore a feature
based grouping of cases may not be possible. In the
experiment, 800 cases are chosen randomly as learn-
ing data, and 300 cases are chosen as testing data. The
learning data are further divided into 4, 5, 6, 7 and 8 Figure 4. The average accuracy of collaborative policies.
 A Fuzzy Integral Based Query Dispatching Model
Figure 5. The mean cost of the collaborative policies.
Bounded-Counsel policy and the Peer-Counsel policy.
Here we did not include the Committee policy in the
experiment because it can be viewed as a special case
of Best-Committee policy.
Some limitations of our experiment include: (1)
since the number of cases is fixed, an increase of the
number of agents will decrease their competence cor-
respondingly, as the result, the experimental accuracy
will decrease with the increasing number of the agents
and (2) a pre-processing of the agents’ competence is
required. On the other hand, the merits of our method
are: (1) the computation time for finding a satisfac-
tory solution is comparatively less than the current ap-
proaches, (2) our three policies can provide alternative
case dispatching methods to users according to their
preference, and (3) our approach can be used to model
feature interaction among cases.
6. Conclusions
In this paper, we have presented our approach of dis-
patching query to different CBR agents. The policies
are based on the concept of case and group competence.
The problem of feature interaction among cases is also
tackled using the fuzzy integral model. Our approach
has been demonstrated empirically with some testing
cases from the travel domain, and the result shows that
our approach is better than the existing ones. Further
research includes a more detail investigation of case
feature interactions, as well as their modeling in dis-
tributed CBR environments. Furthermore, a more the-
oretical analysis and evaluation of our approach will be
carried out.
309
Acknowledgment
This project is supported by the Hong Kong Poly-
technic University research grant H-ZJ90 and CERG
research grant B-Q379.
References
1. M.V.N. Prasad, V. Lesser, and S. Lander, “Retrieval and reasoning
in distributed case-bases,” Journal of Visual Communication and
Image Representation, Special Issue on Digital Libraries, vol. 7,
no. 1, pp. 74–87, 1996.
2. E. Plaza, J.L. Arcos, and F.J. Martı́n, “Cooperative case-based
reasoning,” in Distributed Artificial Intelligence Meets Machine
Learning: Learning in Multi-Agent Environments, edited by G.
Weiss, 1997, pp. 180–201.
3. D.B. Leake and R. Sooriamurthi, “When two cases are better
than one: Exploiting multiple case-bases,” in Proc. of the 4th Int.
Conf. on Case-Based Reasoning, ICCBR 2001, Vancouver, BC,
Canada, 2001, pp. 321–335.
4. E. Plaza and S. Ontañón, “Ensemble Case-Based Reasoning: Col-
laboration policies for multi-agent cooperative CBR,” in Proc.
of the 4th Int. Conf. on Case-Based Reasoning, ICCBR 2001,
Vancouver, BC, Canada, 2001, pp. 437–451.
5. B. Smyth and E. McKenna, “Modeling the competence of case-
bases,” in Proc. of the 4th European Workshop, EWCBR 1998,
Dublin, Ireland, 1998, pp. 208–220.
6. X.Z. Wang and D.S. Yeung, “Using fuzzy integral to modeling
case-based reasoning with feature interaction,” IEEE Int. Conf.
on Systems, Man, and Cybernetics, vol. 5, pp. 3660–3665, 2000.
7. S.C.K. Shiu, Y. Li, and X.Z. Wang, “Using fuzzy in-
tegral to model case-base competence,” in Proc. of Soft
Computing in Case-Based Reasoning Workshop, in conjunc-
tion with the 4th Int. Conf. in Case-Based Reasoning, IC-
CBR 2001, Vancouver, Canada, 2001, pp. 206–212. Or
available from http://www.aic.nrl.navy.mil/papers/2001/AIC-01-
003/ws5/ws5toc6.pdf, 2002.
8. Z.Y. Wang and G.J. Klir, Fuzzy Measure Theory, Plenum: New
York, USA, 1992, pp. 42–43.
Simon C.K. Shiu is an Assistant Professor at the Department of
Computing, Hong Kong Polytechnic University, Hong Kong. He
received M.Sc. degree in Computing Science from University of
Newcastle Upon Tyne, U.K. in 1985, M.Sc. degree in Business Sys-
tems Analysis and Design from City University, London in 1986
and Ph.D. degree in Computing in 1997 from Hong Kong Polytech-
nic University. He worked as a system analyst and project manager
310 Shiu, Li and Zhang
between 1985 and 1990 in several business organizations in Hong
Kong. His current research interests include Case-base Reasoning,
Machine Learning and Soft Computing. He has co-authored (with
Professor Sankar K. Pal) a research monograph Foundations of Soft
Case-Based Reasoning published by John Wiley in 2004.
Dr. Shiu is a member of the British Computer Society and the
IEEE.
Yan Li received the B.Sc. and M.Sc. degrees in Mathematics in 1998
and 2001 respectively from the College of Computer and Mathemat-
ics, Hebei University, P.R. China. Currently, she is a Ph.D. student in
the Department of Computing, the Hong Kong Polytechnic Univer-
sity. Her interests include fuzzy mathematics, case-based reasoning,
rough sets theory and information retrieval.
Feng Zhang received the B.Sc. and M.Sc. degrees in Mathemat-
ics in 1998 and 2001 respectively from the College of Computer
and Mathematics, Hebei University, P.R. China. Currently, she is a
Lecturer in the College of Computer and Mathematics, Hebei Uni-
versity. Her interests include fuzzy mathematics, neural networks,
case-based reasoning and information retrieval.