- Email: [email protected]
An integrated knowledge-based system for urban planning decision support
Knotoledge-Based 5VSTEFIff'--ELSEVIER
Knowledge-Based Systems I 0 (1997) 103-109
An integrated knowledge-based system for urban planning decision support F e n g S h a n a, Li D. X u b'* ~lnstitute of Systems Engineering, Huazhong Universi(v of Science and Technology, Wuhan 430074, China bDepartment of Management Science and Information Systems, Wright State Universi(v, Dayton. OH 45435, USA Received 18 April 1996; revised 16 January 1997; accepted 30 January' 1997
Abstract More applications that integrate knowledge-based decision support systems and artificial neural networks are starting to appear, and interest in such integrated systems is growing rapidly. The paper presents an integrated system in which a knowledge-based decision support system is integrated with a multilayer artificial neural network for urban planning. By integrating decision support systems, knowledge-based systems and artificial neural networks, the system achieves improvements in the implementation of each as well as increases the scope of the application. This approach is very rewarding in its synergism of three technologies to solve complex problems. The paper discusses the structure of the integrated system, as well as providing an example of application. © 1997 Elsevier Science B.V.
Keywords: Knowledge-based systems; Decision support systems; Artificial neural networks; Urban planning
1. Introduction Hybrid architectures for intelligent systems is a new field of artificial intelligence research concerned with the development of the next generation of intelligent systems. Currently a synergism is rapidly developing in the fields of expert systems and neural networks, and an understanding is starting to develop about the theoretical basis and methodology for integrating these two technologies. The research interests in the fields focus on integrating the computational paradigms of expert systems and neural networks, both conventional and fuzzy, and exploring the underlying structures of these two methods of knowledge manipulation, as well as on various applications in which intelligent hybrid systems may and can play an important role. Neural networks can analyze large quantities of data to establish patterns and characteristics in situations where rules are not known and can in many cases make sense of incomplete or noisy data. These capabilities have thus far proven difficult for the traditional symbolic/logic approach. The complementarity of neural networks and expert systems make hybrid systems a very promising area for research
* Corresponding author. Tel.: +l 937 775 2890; fax: +1 937 429 9568: e-maih Ixu @alpha.wright.edu 0950-7051/97/$17.00 © 1997 Elsevier Science B.V. All rights reserved PII S 0 9 5 0 - 7 0 5 1 ( 9 7 ) 0 0 0 0 6 - 3
and development. Recently more applications that integrate knowledge-based decision support systems and artificial neural networks (ANN) are starting to appear, and interest in such hybrid systems is growing rapidly [1,2]. In this paper, we present a knowledge-based decision support system that integrates with a multilayer ANN for urban planning. In the last decade, as a result of economic reform, China has experienced significant structural changes in the national economy. The decision-makers (DMs) in the Chinese public sector are eager to trace the pace as well as the trends of such changes. Since large cities are one of the focuses of structural changes, DMs pay close attention to the comprehensive evaluation of the development of large cities. Such comprehensive evaluation processes are often highly complex and require voluminous input data to be mapped through a substantial number of logical and quantitative interactions; and it is expected that the evaluation will provide insights, into the cause-effect relationship between a variety of factors such as natural, financial resources and the level of development. The resultant evaluation is expected to provide useful decision support to the policy-makers. As mentioned earlier, the integration of expert systems and ANNs is an ideal step in developing intelligent systems since the two methods complement each other such that expert systems allow hard constraints, while ANNs accommodate soft constraints. Specifically,
104
F. Shah, L.D. Xu/Knowledge-Based 5~vstems 10 (1997) 103 109
expert systems involve formal logic and rule interpretation, while ANNs perform nonlinear functions and pattern recognition capabilities. In this study, we explore their complementary strengths to create a hybrid system for urban planning. First, the comprehensive evaluation of urban development is conducted by a group of experts from various fields and a number of approaches are used to aggregate individual opinions into a group consensus [3]. Based on the idea of decision support system (DSS)/knowledge-based system (KBS) integration, a knowledge-based decision support system for comprehensive evaluation of urban development (KB-CEDSS) is constructed as a front-end to the ANN [4,5]. The main function of the KB-CEDSS is to elicit and organize expert opinions, to display analytical results and to demonstrate policy alternatives. For each city being evaluated, the KB-CEDSS generates a set of observation pairs Ix, fix)], i.e. factor index and evaluation result indicator. Second, the [x, fix)] pairs generated by the KB-CEDSS are used as training samples as well as a validation set to train the multilayer ANN, such that we complement the knowledge-based evaluation conducted by the KB-CEDSS with black box models of ANN. It is expected that a welltrained ANN can rapidly process input vectors to produce associated facts and results for the evaluation task. After supervised training processes, the ANN abstracts and generalizes the information provided by [x, fix)] pairs, and produces an output vector. Third, current maintenance of the knowledge base of a KBS is mostly done manually. In complex decision environments, expert knowledge is limited from time to time and the knowledge base needs to be refined continuously. ANNs can be used as a knowledge refinement paradigm. The ANNs, due to their pattern recognition characteristics, support the implementation of automated knowledge refinement. In this study, we use the output from the ANN to facilitate the automation of knowledge base maintenance. The recursive process follows: the KB-CEDSS knowledge-based comprehensive evaluation supervises the training of the ANN, and the output of the ANN automatically refines the knowledge of the KB-CEDSS (the refinement of those imprecise and incomplete rules which were obtained initially). Fourth, the integrated system is based on a knowledgebased DSS that incorporates techniques from approximate reasoning in conjunction with a neural network model. It is expected that the integration of DSS, KBS and ANN will have the potential to provide solutions that no single system alone can deliver. The paper is organized as follows: Section 2 provides a description of the KB-CEDSS, Section 3 presents a description of the ANN, Section 4 provides application examples to demonstrate the usefulness of the system, while Section 5 presents the conclusions and discusses future research.
2. Knowledge-based decision support systems (KB-CEDSS) 2.1. System architecture The KB-CEDSS is a complex system consisting of a number of individual subsystems. The framework of KB-CEDSS can be represented as, SKB - CEDSS= (IKB- CEOSS,D, D *, M, M *, A, A *, R, G)
(1) IKB - CEDSS =
(1, KB)
(2)
KB = ( O B J E C T - KB, A N A L Y S I S - KB, T O O L S - KB)
(3) D = (DBO, DBI)
(4)
where SKB-CEDSSdenotes the system KB-CEDSS, in which the IKB-CEDSS represents the dialogue management subsystem of the KB-CEDSS and includes two components: the interface (/) and the knowledge base (KB). KB is subdivided into three components: OBJECT-KB provides knowledge on " w h a t " kind of evaluation index system should be used when different objects are evaluated (a comprehensive evaluation index system is shown in Fig. l [6]), ANAL YSIS-KB provides knowledge on " h o w " to make evaluation on the objects given, and TOOLS-KB provides knowledge on " h o w " to use the tools in the system. D represents the database and includes two subsystems: the Database for Objects (DBO) stores the attributes of the original indexes of the objects to be evaluated, and the Database for Indexes (DB1) stores the structured index attributes according to the index formed and the relevant membership functions. D* represents the database management system (DBMS); M represents the set of mathematical models that can be used to evaluate urban development; M* represents the model management system; A represents the set of programs; A* represents the program management system; R represents various types of report generators; and G represents visual display capabilities. 2.2. Fuzzy information processing Human elements have played a crucial role in the evaluation process. In addition, the evaluation of the overall development of large cities consists of a multi-participant setting. The participants include individuals, small groups as well as large organizations. When such a variety of parties are involved, their preferences and value systems are often diverse since an area of expertise can be viewed diversely by different experts. Therefore, issues related to vagueness, imprecision and ambiguity in human judgments should find a proper place in the formal evaluation process.
105
F. Shan, L.D. Xu/Knowledge-Based Systems 10 (1997) 103-109
Traditional study of such issues is conducted using probabilistic tools and techniques [7]. However, it is not difficult to see that aspects related to imprecision or vagueness clearly have a non-probabilistic character since they are related to imprecision of meanings. Thus, a proper tool for their analysis seems to be the fuzzy set theory and its related possibility theory that makes it possible to formally represent imprecise concepts. Therefore, an important issue in the development of automated decision aids for urban development is handling fuzziness since the evaluations involve human expertise and knowledge, which are invariably imprecise or incomplete. This would enable the system to better emulate human evaluation processes. Several experiments in KBSs address the problem of developing approximate reasoning methods for dealing with imprecise data [8,9]. In our approach, we use a fuzzy logic framework that provides an appropriate language for both acquiring and representing the fuzzy components underlying experts' knowledge. In the KB-CEDSS, conventional mathematical tools and fuzzy mathematical tools are organized in parallel within TOOLS-KB. Linguistic terms involving indexes, weights, the antecedent and consequent parts of rules are encoded as fuzzy sets. The system is able to represent experts' knowledge using membership functions and fuzzy production rules [10].
General Criteria
Maerolayer Criteria
Midlayer Criteria
Mierolayer Criteria T
V -
-
011
F- u l l 1,tl4
--
01
012 - - - ~ ---013
[
- - - -
021
F- - 1 2 4 l ---U48
02
022 ~ _ _ ~ - - - - ~ 5 1 ---U53 -
--
--1"t3l ----1132
- -
!' 0 - -
--bl21 b128
-
03
023--~
--lt6] -----U63 ~- - u 7 1 ..... U75
Fig. 1. The four-layer c o m p r e h e n s i v e e~,aluation index system.
judgment and object description. For example, the verbal expression of weights such as " a little bit stronger than ..." and a measure of "about 30%" can be represented. Those fuzzy operations and rules are packed into an independent module as a component of TOOLS-KB [ 11]. 2.4. Structuring knowledge bases
2.3. M3CEP
The knowledge base KB = (OBJECT-KB, ANALYSIS-KB, TOOLS-KB) is of large scale. For example, OBJECT-KB
The comprehensive evaluation on large cities made by experts can be formalized as a multiparticipant, multilayer, multicriteria evaluation problem M3CEP and represented as follows [6],
has four layers that are consistent with the layer objectives. Structuring is essential in order to manage such a knowledge base. Since the desired knowledge support depends on changing evaluation environments, i.e. the changing E M in (5), a structured, modular knowledge base is always defined with respect to a current evaluation focus. Such structuring has ameliorated the problems in efforts to make a knowledge base comprehensible and maintainable. There are two kinds of module connections. An ANDconnection implies that modules on the top layer can input and do not contain competing knowledge, but rather contain information on different topics of the related domain; in other words, all of these knowledge may be used at the same time. The second, OR-connection implies that the modules on the lower layer contain competing knowledge; in this case, only one of these modules may be used at a given time. The combination of modules is implemented in such a way that lower layer modules' results can be the input of a corresponding upper layer. This is an experiment on the organization of intelligence that mimics the actual behavior of human experts. Namely when the problem locus changes, the knowledge-base will correspondingly be reorganized.
M3CEP = ( C, IS, EM, EO, ER, E)
(5)
In (5), the evaluative object set, i.e. cities C = (C~, C2 . . . . . Ck . . . . . Cm), 1 < k < m; IS represents the index system of hierarchical architecture (see Fig. 1), where rk,ht represents expert t's evaluation of the uth index in the kth layer t t t t of the hth object in the set C, and rkuh = rk, l, rku2. . . . . rk,~ . . . . . rtkum; E M represents the time-area-event indicator, for example, the time as year 1996, the area as nine large cities, and the event as annual regular evaluation; EO represents the current evaluation goal, e.g. the spacial structure of investment decisions; ER represents aggregation, e.g. GER1 represents an aggregation of individual expert's C3CEP final results, and GES2 indicates that an indexspecific aggregation is completed and reduced to the overall measure; and E represents human expert's set E = (Et, E2 . . . . . Et). Given C, IS, EM, and EO, a matrix B = (b j, bz ..... b,,) is generated by E according to ER. For each city being evaluated, a mapping xt ---' b / i s validated and called the observation-result pair or [x, f(x)] = b pair. Fuzzy techniques are applied in the process to solve (5) in such a way that allows fuzzy measures to be presented in both subjective
3. Artificial neural n e t w o r k s
One of the primary attractions of the ANN approach is
F. Shan, L.D. Xu/Knowledge-Based Systems 10 (1997) 103-109
106
Input Layer
1st Hidden Layer
2nd Hidden Layer
Output Layer
2
rt
Fig. 2. The four-layer ANN.
that knowledge is ascertained directly from accumulated case data through the use of a learning algorithm that may be either supervised or unsupervised. The advantages of ANN include the ability to classify patterns that vary in an unknown manner, recognize patterns within noise, and recall patterns even if some processing units fail [12]. However, ANNs fall short where KBSs excel; such as handling logic, heuristics, and domain knowledge. Therefore, expert systems and ANNs present complementary approaches to knowledge representation: the logical, cognitive, and mechanical nature of the expert systems versus the numeric, associative, and self-organizing nature of the neural network. Compared to conventional methods, the ANN approach shows particular promise in domains where features of different types collectively contribute to the solution of a problem such as a multi-participant, multicriteria decision-making problem. A typical example of such a problem is the comprehensive evaluation of urban development. By combining the powers of expert systems and neural networks in an hybrid system, one that allows for imprecise information and/or uncertain environments, we would have a system more powerful than either one of its components standing alone. In order to benefit from the capabilities of each method, this study uses a hybrid architecture that integrates KBSs and ANNs to generate solutions for urban planning.
described as a function of X and the architecture parameter p, F = N(p, X)
and the parameter adjusting process can be expressed with the learning function as dp(t) = L(p( t), Xi, Yi) dt
(7)
where Xi, Yi are the ith sample pair. For simplicity, write Y
=fix). Practical experiences show that the capabilities of 2-hidden layered perceptrons often exceed the capabilities of l- or 0-hidden layered perceptrons. However, more than 2 hidden layers do not help [13]. We initialize the structure of the ANN as a 4-layer neural network for evaluating urban development (see Fig. 2). For the comprehensive evaluation of the development level of a group of cities, n input and m output units are chosen to match the [x,f(x)] pairs specified by the evaluation goal. The number of units in the first and second hidden layer is determined by the ongoing learning procedure. The self-configuration algorithm is adopted. The algorithm defines a factor r 0 that measures the strength of the relationship between unit i and unitj within the same hidden layer,
3.1. Learning in neurocomputing ANNs are useful because they can be considered as a way of learning knowledge without prior specification of a representation scheme. Utilizing the ANN characteristics of content-based retrieval, approximate matching and simple relaxation and learning, we use the [x, fix)] pairs generated by B as the sample sets to train the multilayer ANN. Learning in neurocomputing is the process of finding or creating the "correct" network architecture, e.g. searching for the best synaptic strengths and threshold values. Given the ANN input vector X, the output vector Y can be
(6)
n = l Yip'Yip -- Yi'Yi n
, y2p _ ~
y2 _ 92 p=l
where Yip is the output value of unit i for the pth sample, and Yi,Yj are defined as the output mean of ij units as follows: Yi = n E
p |=
Yip, Yj =
p=l
Yjp
(9)
and n denotes the number of samples included in the training set. We define the scatter degree parameter as
F. Shan, L.D. Xu/Knowledge-BasedSystems 10 (1997) 103-109
In the second method, we use momentum terms as follows:
follows: 1
107
n
si = -~~
wij(n + 1) = w(n) - ~(n + 1)z(n)
"~
y~,, - ~-,~
(16)
( l o)
p=l
The square root of Si is also present in the denominator of r~j. We then apply the following two rules: Rule 1. If Ir~jl > C1, and S~, Sj > Ce, then combine unit i and j to be unit i. Rule 2. If Sj < C2, then unit j can be deleted. Using the self-configuration algorithm, C1 E [.8,.9], C: E [.001,.01 ]. It helps determine the number of units during the learning process.
3.2. Fast backpropagation algorithm In implementing the BP algorithm for the multilayer feedforward ANN, for each input-output pair [x, fix)], a forward pass starting at the input units computes the activity level yi of all the units in the network. Then a backward pass starting at the output units computes OE[Oyi for each unitj of hidden layer J. Suppose there is another linked layer K next to J, we write:
OE ~ O E d y ~ d x k c3~jj= . Oyk dx~ dyj -
~OEdy k ~yk --~kWkj
(11)
For adjusting the back connection strength Wij, the following equation applies: 0E aw, j
Wij = - r / - - -
(12)
z(n) :
OE t- az(n - 1) cgwij(n )
(17)
where c~ is a constant chosen from [0,11.
3.3. Variable-shaped activation An important aspect to improve the learning performance of ANNs is to apply the property of activation function. Instead of using the sigmoid function, i.e. fix) = I/(1 + e -h) in the standard BP, we can use the function g(x) = Af(Bx) + C and view it as a variable-shaped extension of the function fix). A, B, C are parameters that promote the magnitude, attitude and starting levels of the sigmoid function and are controllable with respect to the given problem to be solved.
4. Applications 4.1. Self-configuration by learning 4.1.1. Task Classify a number of Chinese cities into three categories (A, B, C) by their statistics of six aspects given the ANN with two hidden layers and units distributed as (6-6-4-3).
4.1,2. Solution where ~ is the learning rate or the iteration stepsize. In a standard BP, r/is a given constant. In the Fast Backpropagation (FBP), there are various methods to make the ~ variable speed up the learning process (examples follow). The first method is adaptive variable stepsize technique. The underlying idea of this algorithm is that if the signs of 8E/Owij for two continuous iterations are different, it is obvious that an "overshoot" takes place and the stepsize should be decreased; if the signs of OE/Owo for two continuous iterations are the same, then it means that the speed of convergence is slow and the stepsize should be increased. A linear reinforcement algorithm is used to adapt the stepsize 7,
Step 1. Check the ANN structure with the sample set, using correlation parameter rij and scatter parameter Si of the self-configuration algorithm, conditional on the standard BP, where r/ = .3, C~ = .95, C2 = .001, and global error E = .01. The resultant data are shown in Table 1 Table 2. The data indicate that in the first hidden layer, units 1, 2 and 4 and also units 5 and 6 should be combined into one. In the second hidden layer, units 2 and 4 should be combined. Thus the resultant ANN structure is (6-3-3-3). Step 2. Utilize the newly determined (6-3-3-3) structure to fulfill the classification task. For any city being evaluated, the associated category is given immediately after the input.
An(n + 1) =/3X~q(n)
Table 1 The first hidden layer
(13)
in which,/3 is a constant of [0,1], n is the number of adjusting times, and X is defined as
i~E OE X = sgn[ Owij-~-+ 1) O w ~ l
(14)
The rule for adapting connection strength is
OE Awij(n + 1) : - rl(n + 1) Owij(n)
(15)
i
1 2 3 4 5 6
j 1
2
3
4
5
6
Si
1.000 0.9714 0.0001 0.9886 0.3139 0.2883
0.9714 1.000 0.2167 0.9664 0.4647 0.6116
0.0010 0.2167 1.000 0.6219 0.4969 0.8352
0.9886 0.9664 0.6219 1.000 0.8433 0.4256
0.3139 0.4647 0.4969 0.8433 1.000 0.9594
0.2883 0.6116 0.8352 0.4256 0.9594 1.000
0.0480 0.0970 0.0040 0.0750 0.0240 0.0170
108
F. Shan, L.D. XufKnowledge-Based @stems 10 ( 1997) 103-109
Table 2 The second hidden layer i
1 2 3 4
Table 4 Comprehensive evaluation results by KB-CEDSS
j 1
2
3
4
Si
1.000 0.8808 0.0539 0.9215
0.8808 1.000 0.2957 0.9706
0.0539 0.2957 1.000 0.1151
0.9815 0.9706 0.1151 1.000
0.0030 0.0370 0.0880 0.0410
4.2. F E B versus B P
Table 3 shows the performance of FBP relative to that of standard BP when o~ = .5,/3 = .25, rl(0) = .3, E = .01 and the maximal iteration number is 4000. 4.3. Valid set and outputs
Table 4 shows comprehensive evaluation results generated by the K B - C E D S S for nine cities. The evaluation uses the 1991 statistics of 33 items in social and environmental aspects as characteristics [14]. The input/output pairs, average of experts' knowledge, are qualified to form a sample set for training the corresponding A N N to replicate the evaluation provided by the KB-CEDSS. With the sample set, a supervised training process conducted by FBP and the self-configuration algorithm has established an A N N structure of (33-9-3-3). Table 5, a summary of the computational results for six cities, provides a comparison of the K B - C E D S S fuzzy comprehensive evaluation results versus the A N N evaluation results. One can see both the differences and commonality of the results upon close observation of Table 5. 4.4. Test and evaluation
During training externally provided data are compared with the neural network outputs, and the feedback is used to adjust the weights until all training patterns are correctly categorized by the network. To evaluate the effectiveness of the hybrid system, we compare the results generated by the K B - C E D S S and A N N with the corresponding results provided by human experts. The test results are as follows: (i) given the set of test data, the results generated by both the K B - C E D S S and A N N are comparable with experts' evaluation; (ii) based on detailed analysis of the features of the K B - C E D S S and A N N as well as the application domain, the differences between the results are considered to be
Shanghai Beijing Tianjin Shenyang Wuhan Guangzhou Harbin Chongqing Nanjing
Standard BP
Parameter
rl = 0.3 ~=0 1541 9.45
Iteration steps BP/FBP
Fast BP r/= 0.3 ~:0.5 579 3.55
rt(0) = 0.3 ~=0.5 163 I
Environmental Comprehensive Order index index
0.663 0.654 0.403 0.435 0.435 0.613 0.377 0.211 0.512
0.422 0.658 0.259 0.508 0.433 0.246 0.487 0.324 0.250
0.635 0.643 0.364 0.390 0.364 0.502 0.328 0.227 0.334
2 1 6 4 5 3 8 9 7
mainly caused by the characteristics of each technology since each of them emphasizes certain aspects of the problem domain. For example, K B - C E D S S is good at providing knowledge-based quantitative analysis while A N N is good at coping with the complexity of urban systems and recognizing patterns in noisy data; and (iii) the commonality between the results have demonstrated that each technology can complement the other to certain extent. In general, the quality of the solution is considered acceptable.
5. Conclusions This paper describes and illustrates an integrated system that combines a knowledge-based DSS and an ANN, with the inclusion of appropriate reasoning. We have found the system to be effective in evaluating urban development. One can apply the general approach utilized in the integrated system to a diverse set of problems in any area of automated decision making. In this study, we have merged three technologies, i.e. DSS, KBS and ANN, in the same complex application. The study confirms the complementary nature of these three technologies. By integrating these three technologies one can achieve improvements in the implementation of each as well as increase the scope of application; therefore, this approach is rewarding in its synergism of three technologies to solve complex problems. The major advantages of the hybrid approach include the following: (i) KBSs are Table 5 KB-CEDSS results versus ANN results
City
KB-CEDSS
Fuzzy comprehensive evaluation
Table 3 FBP versus BP Algorithm
Social index
Tianjin Harbin Xian Zhengzhou Shijiazhuang Tangshan
Trained ANN evaluation
Level
Order
Level
Order
0.334 0.328 0.277 0.290 0.320 0.255
1 2 5 4 3 6
0.349 0.316 0.271 0.283 0.328 0.273
1 3 6 4 2
F. Shan, L.D. Xu/Knowledge-Based Systems 10 (1997) 103-109
good for closed-system applications for which inputs are precise, leading to logical outputs. For applications with well-defined rules, KBSs can provide good performance. A N N s can analyze large quantities of data to establish patterns and characteristics in situations where rules are not known. The hybrid system is able to complement the evaluation provided by the K B - C E D S S using rules with pattern recognition capability of ANNs; (ii) by integrating A N N with KBS we can automate knowledge refinement. The ability to learn in unknown environments is an essential component of any intelligent system and is particularly crucial to its performance. This ability can be enhanced by incorporating neural network learning mechanisms into KBSs, A N N techniques enable the KBS to modify and/or enrich its knowledge structures autonomously. Rules and facts may be frequently modified, and knowledge in rules may be evolutionary, dependent on human experience in the domain. The integrated system offers the means to overcome some of the major drawbacks of conventional KBSs, such as their reliance on consultation with human experts for knowledge refinement, and their inability to synthesize new knowledge. The A N N in the integrated system analyzes the data sets originally derived from experts to identify underlying patterns and relationships that subsequently refine the knowledge of the KBS and produces specific knowledge relevant to the evaluation of urban development. Then, the KBS can perform further analysis. For complex applications, it is obvious that hybrid approaches that combine methods of traditional DSS, KBS and A N N s are more appropriate. Currently, the system is being improved from a number of viewpoints. First, a training database is being set up in the K B - C E D S S environment. The training database could be used to train A N N s with respect to specific tasks in urban planning. Second, it is known that the three technologies complement each other to achieve better coverage of perspectives involved in complex decision-making. More research is required to answer an important question, that is, how to achieve best coverage of perspectives through combining and refining three technologies as well as to avoid their disadvantages by enhancing the desirable properties of each other. Third, the A N N in the system is unable to generalize existing evaluations to related new evaluations, i.e. creative abilities have not yet been well established as indicated by Rumelhart [15]. Research is
109
well underway to establish creativity in the integrated system.
Acknowledgements The authors would like to thank Cai Jun and Zhai Fan for their participation in the development and implementation of the system. The authors are also grateful to the referees for their valuable comments on this article.
References [1] L. Medsker, Hybrid neural network and expert systems, Kluwer, Dordrecht, 1994. [2] L. Medsker, E. Turban, Integrating expert systems and neural computing for decision support, Expert Systems with Applications 7 (1994) 553-562. [3] J. Warfield, A Science of Generic Design: Managing Complexity Through Systems Design, Iowa State University Press, Ames, 1994. [4] M.W. Aiken, O.R. Liu Sheng, D.R. Vogel, Integrating expert systems with group decision support systems, ACM Transactions on Information Systems 9 (1991) 75-95. [5] M.K. E1-Najsawi,A.C. Stylianou, Expert support systems: integrating AI technologies, Communications of the ACM 36 (1993) 54-65. [6] S. Feng, L.D. Xu, Decision support for fuzzy comprehensive evaluation of urban development, Fuzzy Sets and Systems, to appear. [7] E. Binaghi, O.D. Giorgi, G. Maggi, T. Motta, A. Rampini, Computerassisted diagnosis of postmenopausal osteoporosis using a fuzzy expert system shell, Computers and Biomedical Research 26 (1993) 498-516. [8] S.K. Pal, Fuzzy tools for the management of uncertainty in pattern recognition, image analysis, vision and expert systems, International Journal of Systems Science 22 (1991) 511-550. [9] L.D. Xu, Case-based reasoning for AIDS initial assessment, Knowledge-based Systems 8 (1995) 32-38. [10] H.J. Zimmermann, Fuzzy Set Theory and Its Applications, Kluwer, Dordrecht, 1985. [11] S. Feng, Modeling topics in intelligent decision support systems, Information and Control 22 (1993) 364-370. [12] A.K. Jain, J.C. Mao, K.M. Mohiuddin, Artificial neural networks, Computer 29 (1996) 31-44. [13] X. Cui, K.G. Shin, Direct control and coordination using neural networks, IEEE Transactions on Systems Man and Cybernetics 23 (1993) 686-697. [ 14] Chinese State Statistical Bureau, Chinese Urban Statistics Year Book, Chinese Statistical Press, Beijing, 1991. [15] D,E. Rumelhart, B. Widrow, M.A. Lehr, The basic ideas in neural networks, Communications of the ACM 37 (1994) 87-92.
Knowledge-Based Systems I 0 (1997) 103-109
An integrated knowledge-based system for urban planning decision support F e n g S h a n a, Li D. X u b'* ~lnstitute of Systems Engineering, Huazhong Universi(v of Science and Technology, Wuhan 430074, China bDepartment of Management Science and Information Systems, Wright State Universi(v, Dayton. OH 45435, USA Received 18 April 1996; revised 16 January 1997; accepted 30 January' 1997
Abstract More applications that integrate knowledge-based decision support systems and artificial neural networks are starting to appear, and interest in such integrated systems is growing rapidly. The paper presents an integrated system in which a knowledge-based decision support system is integrated with a multilayer artificial neural network for urban planning. By integrating decision support systems, knowledge-based systems and artificial neural networks, the system achieves improvements in the implementation of each as well as increases the scope of the application. This approach is very rewarding in its synergism of three technologies to solve complex problems. The paper discusses the structure of the integrated system, as well as providing an example of application. © 1997 Elsevier Science B.V.
Keywords: Knowledge-based systems; Decision support systems; Artificial neural networks; Urban planning
1. Introduction Hybrid architectures for intelligent systems is a new field of artificial intelligence research concerned with the development of the next generation of intelligent systems. Currently a synergism is rapidly developing in the fields of expert systems and neural networks, and an understanding is starting to develop about the theoretical basis and methodology for integrating these two technologies. The research interests in the fields focus on integrating the computational paradigms of expert systems and neural networks, both conventional and fuzzy, and exploring the underlying structures of these two methods of knowledge manipulation, as well as on various applications in which intelligent hybrid systems may and can play an important role. Neural networks can analyze large quantities of data to establish patterns and characteristics in situations where rules are not known and can in many cases make sense of incomplete or noisy data. These capabilities have thus far proven difficult for the traditional symbolic/logic approach. The complementarity of neural networks and expert systems make hybrid systems a very promising area for research
* Corresponding author. Tel.: +l 937 775 2890; fax: +1 937 429 9568: e-maih Ixu @alpha.wright.edu 0950-7051/97/$17.00 © 1997 Elsevier Science B.V. All rights reserved PII S 0 9 5 0 - 7 0 5 1 ( 9 7 ) 0 0 0 0 6 - 3
and development. Recently more applications that integrate knowledge-based decision support systems and artificial neural networks (ANN) are starting to appear, and interest in such hybrid systems is growing rapidly [1,2]. In this paper, we present a knowledge-based decision support system that integrates with a multilayer ANN for urban planning. In the last decade, as a result of economic reform, China has experienced significant structural changes in the national economy. The decision-makers (DMs) in the Chinese public sector are eager to trace the pace as well as the trends of such changes. Since large cities are one of the focuses of structural changes, DMs pay close attention to the comprehensive evaluation of the development of large cities. Such comprehensive evaluation processes are often highly complex and require voluminous input data to be mapped through a substantial number of logical and quantitative interactions; and it is expected that the evaluation will provide insights, into the cause-effect relationship between a variety of factors such as natural, financial resources and the level of development. The resultant evaluation is expected to provide useful decision support to the policy-makers. As mentioned earlier, the integration of expert systems and ANNs is an ideal step in developing intelligent systems since the two methods complement each other such that expert systems allow hard constraints, while ANNs accommodate soft constraints. Specifically,
104
F. Shah, L.D. Xu/Knowledge-Based 5~vstems 10 (1997) 103 109
expert systems involve formal logic and rule interpretation, while ANNs perform nonlinear functions and pattern recognition capabilities. In this study, we explore their complementary strengths to create a hybrid system for urban planning. First, the comprehensive evaluation of urban development is conducted by a group of experts from various fields and a number of approaches are used to aggregate individual opinions into a group consensus [3]. Based on the idea of decision support system (DSS)/knowledge-based system (KBS) integration, a knowledge-based decision support system for comprehensive evaluation of urban development (KB-CEDSS) is constructed as a front-end to the ANN [4,5]. The main function of the KB-CEDSS is to elicit and organize expert opinions, to display analytical results and to demonstrate policy alternatives. For each city being evaluated, the KB-CEDSS generates a set of observation pairs Ix, fix)], i.e. factor index and evaluation result indicator. Second, the [x, fix)] pairs generated by the KB-CEDSS are used as training samples as well as a validation set to train the multilayer ANN, such that we complement the knowledge-based evaluation conducted by the KB-CEDSS with black box models of ANN. It is expected that a welltrained ANN can rapidly process input vectors to produce associated facts and results for the evaluation task. After supervised training processes, the ANN abstracts and generalizes the information provided by [x, fix)] pairs, and produces an output vector. Third, current maintenance of the knowledge base of a KBS is mostly done manually. In complex decision environments, expert knowledge is limited from time to time and the knowledge base needs to be refined continuously. ANNs can be used as a knowledge refinement paradigm. The ANNs, due to their pattern recognition characteristics, support the implementation of automated knowledge refinement. In this study, we use the output from the ANN to facilitate the automation of knowledge base maintenance. The recursive process follows: the KB-CEDSS knowledge-based comprehensive evaluation supervises the training of the ANN, and the output of the ANN automatically refines the knowledge of the KB-CEDSS (the refinement of those imprecise and incomplete rules which were obtained initially). Fourth, the integrated system is based on a knowledgebased DSS that incorporates techniques from approximate reasoning in conjunction with a neural network model. It is expected that the integration of DSS, KBS and ANN will have the potential to provide solutions that no single system alone can deliver. The paper is organized as follows: Section 2 provides a description of the KB-CEDSS, Section 3 presents a description of the ANN, Section 4 provides application examples to demonstrate the usefulness of the system, while Section 5 presents the conclusions and discusses future research.
2. Knowledge-based decision support systems (KB-CEDSS) 2.1. System architecture The KB-CEDSS is a complex system consisting of a number of individual subsystems. The framework of KB-CEDSS can be represented as, SKB - CEDSS= (IKB- CEOSS,D, D *, M, M *, A, A *, R, G)
(1) IKB - CEDSS =
(1, KB)
(2)
KB = ( O B J E C T - KB, A N A L Y S I S - KB, T O O L S - KB)
(3) D = (DBO, DBI)
(4)
where SKB-CEDSSdenotes the system KB-CEDSS, in which the IKB-CEDSS represents the dialogue management subsystem of the KB-CEDSS and includes two components: the interface (/) and the knowledge base (KB). KB is subdivided into three components: OBJECT-KB provides knowledge on " w h a t " kind of evaluation index system should be used when different objects are evaluated (a comprehensive evaluation index system is shown in Fig. l [6]), ANAL YSIS-KB provides knowledge on " h o w " to make evaluation on the objects given, and TOOLS-KB provides knowledge on " h o w " to use the tools in the system. D represents the database and includes two subsystems: the Database for Objects (DBO) stores the attributes of the original indexes of the objects to be evaluated, and the Database for Indexes (DB1) stores the structured index attributes according to the index formed and the relevant membership functions. D* represents the database management system (DBMS); M represents the set of mathematical models that can be used to evaluate urban development; M* represents the model management system; A represents the set of programs; A* represents the program management system; R represents various types of report generators; and G represents visual display capabilities. 2.2. Fuzzy information processing Human elements have played a crucial role in the evaluation process. In addition, the evaluation of the overall development of large cities consists of a multi-participant setting. The participants include individuals, small groups as well as large organizations. When such a variety of parties are involved, their preferences and value systems are often diverse since an area of expertise can be viewed diversely by different experts. Therefore, issues related to vagueness, imprecision and ambiguity in human judgments should find a proper place in the formal evaluation process.
105
F. Shan, L.D. Xu/Knowledge-Based Systems 10 (1997) 103-109
Traditional study of such issues is conducted using probabilistic tools and techniques [7]. However, it is not difficult to see that aspects related to imprecision or vagueness clearly have a non-probabilistic character since they are related to imprecision of meanings. Thus, a proper tool for their analysis seems to be the fuzzy set theory and its related possibility theory that makes it possible to formally represent imprecise concepts. Therefore, an important issue in the development of automated decision aids for urban development is handling fuzziness since the evaluations involve human expertise and knowledge, which are invariably imprecise or incomplete. This would enable the system to better emulate human evaluation processes. Several experiments in KBSs address the problem of developing approximate reasoning methods for dealing with imprecise data [8,9]. In our approach, we use a fuzzy logic framework that provides an appropriate language for both acquiring and representing the fuzzy components underlying experts' knowledge. In the KB-CEDSS, conventional mathematical tools and fuzzy mathematical tools are organized in parallel within TOOLS-KB. Linguistic terms involving indexes, weights, the antecedent and consequent parts of rules are encoded as fuzzy sets. The system is able to represent experts' knowledge using membership functions and fuzzy production rules [10].
General Criteria
Maerolayer Criteria
Midlayer Criteria
Mierolayer Criteria T
V -
-
011
F- u l l 1,tl4
--
01
012 - - - ~ ---013
[
- - - -
021
F- - 1 2 4 l ---U48
02
022 ~ _ _ ~ - - - - ~ 5 1 ---U53 -
--
--1"t3l ----1132
- -
!' 0 - -
--bl21 b128
-
03
023--~
--lt6] -----U63 ~- - u 7 1 ..... U75
Fig. 1. The four-layer c o m p r e h e n s i v e e~,aluation index system.
judgment and object description. For example, the verbal expression of weights such as " a little bit stronger than ..." and a measure of "about 30%" can be represented. Those fuzzy operations and rules are packed into an independent module as a component of TOOLS-KB [ 11]. 2.4. Structuring knowledge bases
2.3. M3CEP
The knowledge base KB = (OBJECT-KB, ANALYSIS-KB, TOOLS-KB) is of large scale. For example, OBJECT-KB
The comprehensive evaluation on large cities made by experts can be formalized as a multiparticipant, multilayer, multicriteria evaluation problem M3CEP and represented as follows [6],
has four layers that are consistent with the layer objectives. Structuring is essential in order to manage such a knowledge base. Since the desired knowledge support depends on changing evaluation environments, i.e. the changing E M in (5), a structured, modular knowledge base is always defined with respect to a current evaluation focus. Such structuring has ameliorated the problems in efforts to make a knowledge base comprehensible and maintainable. There are two kinds of module connections. An ANDconnection implies that modules on the top layer can input and do not contain competing knowledge, but rather contain information on different topics of the related domain; in other words, all of these knowledge may be used at the same time. The second, OR-connection implies that the modules on the lower layer contain competing knowledge; in this case, only one of these modules may be used at a given time. The combination of modules is implemented in such a way that lower layer modules' results can be the input of a corresponding upper layer. This is an experiment on the organization of intelligence that mimics the actual behavior of human experts. Namely when the problem locus changes, the knowledge-base will correspondingly be reorganized.
M3CEP = ( C, IS, EM, EO, ER, E)
(5)
In (5), the evaluative object set, i.e. cities C = (C~, C2 . . . . . Ck . . . . . Cm), 1 < k < m; IS represents the index system of hierarchical architecture (see Fig. 1), where rk,ht represents expert t's evaluation of the uth index in the kth layer t t t t of the hth object in the set C, and rkuh = rk, l, rku2. . . . . rk,~ . . . . . rtkum; E M represents the time-area-event indicator, for example, the time as year 1996, the area as nine large cities, and the event as annual regular evaluation; EO represents the current evaluation goal, e.g. the spacial structure of investment decisions; ER represents aggregation, e.g. GER1 represents an aggregation of individual expert's C3CEP final results, and GES2 indicates that an indexspecific aggregation is completed and reduced to the overall measure; and E represents human expert's set E = (Et, E2 . . . . . Et). Given C, IS, EM, and EO, a matrix B = (b j, bz ..... b,,) is generated by E according to ER. For each city being evaluated, a mapping xt ---' b / i s validated and called the observation-result pair or [x, f(x)] = b pair. Fuzzy techniques are applied in the process to solve (5) in such a way that allows fuzzy measures to be presented in both subjective
3. Artificial neural n e t w o r k s
One of the primary attractions of the ANN approach is
F. Shan, L.D. Xu/Knowledge-Based Systems 10 (1997) 103-109
106
Input Layer
1st Hidden Layer
2nd Hidden Layer
Output Layer
2
rt
Fig. 2. The four-layer ANN.
that knowledge is ascertained directly from accumulated case data through the use of a learning algorithm that may be either supervised or unsupervised. The advantages of ANN include the ability to classify patterns that vary in an unknown manner, recognize patterns within noise, and recall patterns even if some processing units fail [12]. However, ANNs fall short where KBSs excel; such as handling logic, heuristics, and domain knowledge. Therefore, expert systems and ANNs present complementary approaches to knowledge representation: the logical, cognitive, and mechanical nature of the expert systems versus the numeric, associative, and self-organizing nature of the neural network. Compared to conventional methods, the ANN approach shows particular promise in domains where features of different types collectively contribute to the solution of a problem such as a multi-participant, multicriteria decision-making problem. A typical example of such a problem is the comprehensive evaluation of urban development. By combining the powers of expert systems and neural networks in an hybrid system, one that allows for imprecise information and/or uncertain environments, we would have a system more powerful than either one of its components standing alone. In order to benefit from the capabilities of each method, this study uses a hybrid architecture that integrates KBSs and ANNs to generate solutions for urban planning.
described as a function of X and the architecture parameter p, F = N(p, X)
and the parameter adjusting process can be expressed with the learning function as dp(t) = L(p( t), Xi, Yi) dt
(7)
where Xi, Yi are the ith sample pair. For simplicity, write Y
=fix). Practical experiences show that the capabilities of 2-hidden layered perceptrons often exceed the capabilities of l- or 0-hidden layered perceptrons. However, more than 2 hidden layers do not help [13]. We initialize the structure of the ANN as a 4-layer neural network for evaluating urban development (see Fig. 2). For the comprehensive evaluation of the development level of a group of cities, n input and m output units are chosen to match the [x,f(x)] pairs specified by the evaluation goal. The number of units in the first and second hidden layer is determined by the ongoing learning procedure. The self-configuration algorithm is adopted. The algorithm defines a factor r 0 that measures the strength of the relationship between unit i and unitj within the same hidden layer,
3.1. Learning in neurocomputing ANNs are useful because they can be considered as a way of learning knowledge without prior specification of a representation scheme. Utilizing the ANN characteristics of content-based retrieval, approximate matching and simple relaxation and learning, we use the [x, fix)] pairs generated by B as the sample sets to train the multilayer ANN. Learning in neurocomputing is the process of finding or creating the "correct" network architecture, e.g. searching for the best synaptic strengths and threshold values. Given the ANN input vector X, the output vector Y can be
(6)
n = l Yip'Yip -- Yi'Yi n
, y2p _ ~
y2 _ 92 p=l
where Yip is the output value of unit i for the pth sample, and Yi,Yj are defined as the output mean of ij units as follows: Yi = n E
p |=
Yip, Yj =
p=l
Yjp
(9)
and n denotes the number of samples included in the training set. We define the scatter degree parameter as
F. Shan, L.D. Xu/Knowledge-BasedSystems 10 (1997) 103-109
In the second method, we use momentum terms as follows:
follows: 1
107
n
si = -~~
wij(n + 1) = w(n) - ~(n + 1)z(n)
"~
y~,, - ~-,~
(16)
( l o)
p=l
The square root of Si is also present in the denominator of r~j. We then apply the following two rules: Rule 1. If Ir~jl > C1, and S~, Sj > Ce, then combine unit i and j to be unit i. Rule 2. If Sj < C2, then unit j can be deleted. Using the self-configuration algorithm, C1 E [.8,.9], C: E [.001,.01 ]. It helps determine the number of units during the learning process.
3.2. Fast backpropagation algorithm In implementing the BP algorithm for the multilayer feedforward ANN, for each input-output pair [x, fix)], a forward pass starting at the input units computes the activity level yi of all the units in the network. Then a backward pass starting at the output units computes OE[Oyi for each unitj of hidden layer J. Suppose there is another linked layer K next to J, we write:
OE ~ O E d y ~ d x k c3~jj= . Oyk dx~ dyj -
~OEdy k ~yk --~kWkj
(11)
For adjusting the back connection strength Wij, the following equation applies: 0E aw, j
Wij = - r / - - -
(12)
z(n) :
OE t- az(n - 1) cgwij(n )
(17)
where c~ is a constant chosen from [0,11.
3.3. Variable-shaped activation An important aspect to improve the learning performance of ANNs is to apply the property of activation function. Instead of using the sigmoid function, i.e. fix) = I/(1 + e -h) in the standard BP, we can use the function g(x) = Af(Bx) + C and view it as a variable-shaped extension of the function fix). A, B, C are parameters that promote the magnitude, attitude and starting levels of the sigmoid function and are controllable with respect to the given problem to be solved.
4. Applications 4.1. Self-configuration by learning 4.1.1. Task Classify a number of Chinese cities into three categories (A, B, C) by their statistics of six aspects given the ANN with two hidden layers and units distributed as (6-6-4-3).
4.1,2. Solution where ~ is the learning rate or the iteration stepsize. In a standard BP, r/is a given constant. In the Fast Backpropagation (FBP), there are various methods to make the ~ variable speed up the learning process (examples follow). The first method is adaptive variable stepsize technique. The underlying idea of this algorithm is that if the signs of 8E/Owij for two continuous iterations are different, it is obvious that an "overshoot" takes place and the stepsize should be decreased; if the signs of OE/Owo for two continuous iterations are the same, then it means that the speed of convergence is slow and the stepsize should be increased. A linear reinforcement algorithm is used to adapt the stepsize 7,
Step 1. Check the ANN structure with the sample set, using correlation parameter rij and scatter parameter Si of the self-configuration algorithm, conditional on the standard BP, where r/ = .3, C~ = .95, C2 = .001, and global error E = .01. The resultant data are shown in Table 1 Table 2. The data indicate that in the first hidden layer, units 1, 2 and 4 and also units 5 and 6 should be combined into one. In the second hidden layer, units 2 and 4 should be combined. Thus the resultant ANN structure is (6-3-3-3). Step 2. Utilize the newly determined (6-3-3-3) structure to fulfill the classification task. For any city being evaluated, the associated category is given immediately after the input.
An(n + 1) =/3X~q(n)
Table 1 The first hidden layer
(13)
in which,/3 is a constant of [0,1], n is the number of adjusting times, and X is defined as
i~E OE X = sgn[ Owij-~-+ 1) O w ~ l
(14)
The rule for adapting connection strength is
OE Awij(n + 1) : - rl(n + 1) Owij(n)
(15)
i
1 2 3 4 5 6
j 1
2
3
4
5
6
Si
1.000 0.9714 0.0001 0.9886 0.3139 0.2883
0.9714 1.000 0.2167 0.9664 0.4647 0.6116
0.0010 0.2167 1.000 0.6219 0.4969 0.8352
0.9886 0.9664 0.6219 1.000 0.8433 0.4256
0.3139 0.4647 0.4969 0.8433 1.000 0.9594
0.2883 0.6116 0.8352 0.4256 0.9594 1.000
0.0480 0.0970 0.0040 0.0750 0.0240 0.0170
108
F. Shan, L.D. XufKnowledge-Based @stems 10 ( 1997) 103-109
Table 2 The second hidden layer i
1 2 3 4
Table 4 Comprehensive evaluation results by KB-CEDSS
j 1
2
3
4
Si
1.000 0.8808 0.0539 0.9215
0.8808 1.000 0.2957 0.9706
0.0539 0.2957 1.000 0.1151
0.9815 0.9706 0.1151 1.000
0.0030 0.0370 0.0880 0.0410
4.2. F E B versus B P
Table 3 shows the performance of FBP relative to that of standard BP when o~ = .5,/3 = .25, rl(0) = .3, E = .01 and the maximal iteration number is 4000. 4.3. Valid set and outputs
Table 4 shows comprehensive evaluation results generated by the K B - C E D S S for nine cities. The evaluation uses the 1991 statistics of 33 items in social and environmental aspects as characteristics [14]. The input/output pairs, average of experts' knowledge, are qualified to form a sample set for training the corresponding A N N to replicate the evaluation provided by the KB-CEDSS. With the sample set, a supervised training process conducted by FBP and the self-configuration algorithm has established an A N N structure of (33-9-3-3). Table 5, a summary of the computational results for six cities, provides a comparison of the K B - C E D S S fuzzy comprehensive evaluation results versus the A N N evaluation results. One can see both the differences and commonality of the results upon close observation of Table 5. 4.4. Test and evaluation
During training externally provided data are compared with the neural network outputs, and the feedback is used to adjust the weights until all training patterns are correctly categorized by the network. To evaluate the effectiveness of the hybrid system, we compare the results generated by the K B - C E D S S and A N N with the corresponding results provided by human experts. The test results are as follows: (i) given the set of test data, the results generated by both the K B - C E D S S and A N N are comparable with experts' evaluation; (ii) based on detailed analysis of the features of the K B - C E D S S and A N N as well as the application domain, the differences between the results are considered to be
Shanghai Beijing Tianjin Shenyang Wuhan Guangzhou Harbin Chongqing Nanjing
Standard BP
Parameter
rl = 0.3 ~=0 1541 9.45
Iteration steps BP/FBP
Fast BP r/= 0.3 ~:0.5 579 3.55
rt(0) = 0.3 ~=0.5 163 I
Environmental Comprehensive Order index index
0.663 0.654 0.403 0.435 0.435 0.613 0.377 0.211 0.512
0.422 0.658 0.259 0.508 0.433 0.246 0.487 0.324 0.250
0.635 0.643 0.364 0.390 0.364 0.502 0.328 0.227 0.334
2 1 6 4 5 3 8 9 7
mainly caused by the characteristics of each technology since each of them emphasizes certain aspects of the problem domain. For example, K B - C E D S S is good at providing knowledge-based quantitative analysis while A N N is good at coping with the complexity of urban systems and recognizing patterns in noisy data; and (iii) the commonality between the results have demonstrated that each technology can complement the other to certain extent. In general, the quality of the solution is considered acceptable.
5. Conclusions This paper describes and illustrates an integrated system that combines a knowledge-based DSS and an ANN, with the inclusion of appropriate reasoning. We have found the system to be effective in evaluating urban development. One can apply the general approach utilized in the integrated system to a diverse set of problems in any area of automated decision making. In this study, we have merged three technologies, i.e. DSS, KBS and ANN, in the same complex application. The study confirms the complementary nature of these three technologies. By integrating these three technologies one can achieve improvements in the implementation of each as well as increase the scope of application; therefore, this approach is rewarding in its synergism of three technologies to solve complex problems. The major advantages of the hybrid approach include the following: (i) KBSs are Table 5 KB-CEDSS results versus ANN results
City
KB-CEDSS
Fuzzy comprehensive evaluation
Table 3 FBP versus BP Algorithm
Social index
Tianjin Harbin Xian Zhengzhou Shijiazhuang Tangshan
Trained ANN evaluation
Level
Order
Level
Order
0.334 0.328 0.277 0.290 0.320 0.255
1 2 5 4 3 6
0.349 0.316 0.271 0.283 0.328 0.273
1 3 6 4 2
F. Shan, L.D. Xu/Knowledge-Based Systems 10 (1997) 103-109
good for closed-system applications for which inputs are precise, leading to logical outputs. For applications with well-defined rules, KBSs can provide good performance. A N N s can analyze large quantities of data to establish patterns and characteristics in situations where rules are not known. The hybrid system is able to complement the evaluation provided by the K B - C E D S S using rules with pattern recognition capability of ANNs; (ii) by integrating A N N with KBS we can automate knowledge refinement. The ability to learn in unknown environments is an essential component of any intelligent system and is particularly crucial to its performance. This ability can be enhanced by incorporating neural network learning mechanisms into KBSs, A N N techniques enable the KBS to modify and/or enrich its knowledge structures autonomously. Rules and facts may be frequently modified, and knowledge in rules may be evolutionary, dependent on human experience in the domain. The integrated system offers the means to overcome some of the major drawbacks of conventional KBSs, such as their reliance on consultation with human experts for knowledge refinement, and their inability to synthesize new knowledge. The A N N in the integrated system analyzes the data sets originally derived from experts to identify underlying patterns and relationships that subsequently refine the knowledge of the KBS and produces specific knowledge relevant to the evaluation of urban development. Then, the KBS can perform further analysis. For complex applications, it is obvious that hybrid approaches that combine methods of traditional DSS, KBS and A N N s are more appropriate. Currently, the system is being improved from a number of viewpoints. First, a training database is being set up in the K B - C E D S S environment. The training database could be used to train A N N s with respect to specific tasks in urban planning. Second, it is known that the three technologies complement each other to achieve better coverage of perspectives involved in complex decision-making. More research is required to answer an important question, that is, how to achieve best coverage of perspectives through combining and refining three technologies as well as to avoid their disadvantages by enhancing the desirable properties of each other. Third, the A N N in the system is unable to generalize existing evaluations to related new evaluations, i.e. creative abilities have not yet been well established as indicated by Rumelhart [15]. Research is
109
well underway to establish creativity in the integrated system.
Acknowledgements The authors would like to thank Cai Jun and Zhai Fan for their participation in the development and implementation of the system. The authors are also grateful to the referees for their valuable comments on this article.
References [1] L. Medsker, Hybrid neural network and expert systems, Kluwer, Dordrecht, 1994. [2] L. Medsker, E. Turban, Integrating expert systems and neural computing for decision support, Expert Systems with Applications 7 (1994) 553-562. [3] J. Warfield, A Science of Generic Design: Managing Complexity Through Systems Design, Iowa State University Press, Ames, 1994. [4] M.W. Aiken, O.R. Liu Sheng, D.R. Vogel, Integrating expert systems with group decision support systems, ACM Transactions on Information Systems 9 (1991) 75-95. [5] M.K. E1-Najsawi,A.C. Stylianou, Expert support systems: integrating AI technologies, Communications of the ACM 36 (1993) 54-65. [6] S. Feng, L.D. Xu, Decision support for fuzzy comprehensive evaluation of urban development, Fuzzy Sets and Systems, to appear. [7] E. Binaghi, O.D. Giorgi, G. Maggi, T. Motta, A. Rampini, Computerassisted diagnosis of postmenopausal osteoporosis using a fuzzy expert system shell, Computers and Biomedical Research 26 (1993) 498-516. [8] S.K. Pal, Fuzzy tools for the management of uncertainty in pattern recognition, image analysis, vision and expert systems, International Journal of Systems Science 22 (1991) 511-550. [9] L.D. Xu, Case-based reasoning for AIDS initial assessment, Knowledge-based Systems 8 (1995) 32-38. [10] H.J. Zimmermann, Fuzzy Set Theory and Its Applications, Kluwer, Dordrecht, 1985. [11] S. Feng, Modeling topics in intelligent decision support systems, Information and Control 22 (1993) 364-370. [12] A.K. Jain, J.C. Mao, K.M. Mohiuddin, Artificial neural networks, Computer 29 (1996) 31-44. [13] X. Cui, K.G. Shin, Direct control and coordination using neural networks, IEEE Transactions on Systems Man and Cybernetics 23 (1993) 686-697. [ 14] Chinese State Statistical Bureau, Chinese Urban Statistics Year Book, Chinese Statistical Press, Beijing, 1991. [15] D,E. Rumelhart, B. Widrow, M.A. Lehr, The basic ideas in neural networks, Communications of the ACM 37 (1994) 87-92.