Вы находитесь на странице: 1из 6

Computers and Electronics in Agriculture 66 (2009) 7075

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture


journal homepage: www.elsevier.com/locate/compag

Farm price prediction using case-based reasoning approachA


case of broiler industry in Taiwan
M.L. Shih a , B.W. Huang b, , Nan-Hsing Chiu c , C. Chiu d , W.Y. Hu b
a
Department of Social Studies Education, National Tai-Tung University, Tai-Tung, Taiwan
b
Department of Applied Economics, National Chung-Hsing University, Taichung, Taiwan
c
Department of Information Management, Ching Yun University, Chungli, Taiwan
d
Department of Information Management, Yuan Ze University, Chungli, Taiwan

a r t i c l e i n f o a b s t r a c t

Article history: Since Taiwan joined the World Trade Organization (WTO) in 2002, pricing decision has become more
Received 8 November 2007 essential to the development of the broiler industry. The effective prediction of broiler prices is essential
Received in revised form from the viewpoint of the agriculture authority and the Poultry Association, thus a more realistic broiler
25 November 2008
price structure can assist the government to manage the national production resources more effectively.
Accepted 8 December 2008
This research proposes a weighted case-based reasoning (CBR) approach to construct a price prediction
model. The genetic algorithm model was adopted to nd out the most suitable feature weights for CBR.
Keywords:
Previous local production data and economic indices, along with information about imported chicken,
Broiler price
Case-based reasoning
were collected to build the prediction model. The experimental results indicated that the proposed CBR
Genetic algorithms approach could exhibit a better prediction performance than the ones exhibited by linear regression,
Price prediction regression tree, and neural nets approaches. The ndings also revealed that broiler prices were mostly
inuenced by the prices of colorful broilers and chicks.
2009 Elsevier B.V. All rights reserved.

1. Introduction Besides the NAIF, the Poultry Association is also extremely con-
cerned with the information on the broiler price because the pricing
Because the broiler industry is one of the most important institution of the broiler price is based on an offer system. Under
livestock sectors in Taiwan, the agriculture authority pays high the offer system, poultry processing plants buy broilers from the
attention to its pricing information. According to the Agricultural growers. If the quoted price of the offer increases, the processing
Statistics Yearbook (2005), the industry had a product value of 11.3 plants will have to incur a higher cost price and their prots will be
billion NT dollars and it is ranked as the third highest among Tai- affected. Thus, the association has a specialized unit that is respon-
wans livestock sector in terms of production volumes. sible for collecting all related price information and announcing
This industry growth, however, was greatly affected by the entry the quotation. In real-world practice, a controversy always exists
of imported chicken into the Taiwanese market after the industrys between the processing plants and the specialized unit in the
deregulation in 2005. For example, the number of chickens that Poultry Association. Consequently, the NAIF is sometimes obliged
were slaughtered during the 19962004 period increased from 160 to intervene in their negotiations. To reduce the tension and to
million to 207 million, but signicantly dropped to 170 million in enhance the mutual collaboration of the two parties, the special-
2005. To help chicken raisers overcome the impact of the deregu- ized unit in the Poultry Association needs to provide an effective
lated entry of imported chicken, the agriculture authority created price quote that is based on the related broiler price information. In
the National Animal Industry Foundation (NAIF). One of its tasks other words, a reasonable quotation that can realistically reect the
was to evaluate the broiler price information as a basis for strategic future market price is essential for the sustainable development of
guidance and operational assistance. The setting up of the agency the Taiwanese broiler industry.
implies that evaluating broiler price information has become an The evaluation of broiler price information as implemented
important issue in Taiwan in the context of liberalization and the through the NAIF or the Poultry Association is generally based on
highly volatile international broiler market. expert opinions rather than on a rigorous analysis using reliable
scientic approaches. In the conventional method, the specialized
unit in the Poultry Association collects information such as the chick
quantity, feedstuff price, the imported chicken quantity and price,
Corresponding author. Tel.: +886 4 22840349x212; fax: +886 4 22860255. and the prices of colorful broiler and related livestock; it would then
E-mail address: bwhuang@dragon.nchu.edu.tw (B.W. Huang). announce the price quotation on a daily basis after a discussion with

0168-1699/$ see front matter 2009 Elsevier B.V. All rights reserved.
doi:10.1016/j.compag.2008.12.005
M.L. Shih et al. / Computers and Electronics in Agriculture 66 (2009) 7075 71

experts. On the other hand, the NAIF would suggest a total broiler
quantity required by the growers by relying on the related farm
price information. Meanwhile, the broiler price information has
become more complicated after Taiwan joined the WTO. To improve
the price evaluation performance, it would be necessary to develop
a price evaluation mechanism that uses analytic approaches and
to provide the prediction results on the basis of the discussion and
evaluation of specialists.
This research proposes the case-based reasoning (CBR) approach
for forecasting broiler prices. This approach is not constrained by Fig. 1. The framework of UCBR and LCBR.
the functional form of the econometric model; besides, its estima-
tion is more exible because it can recall previous data patterns.
used cost prediction model. In another study, Oh and Kim (2007)
The uses of the CBR approach have been evaluated and con-
further adopted the CBR approach for nancial market monitor-
rmed in many studies (Bichindaritz, 2006; Chiu and Huang, 2007).
ing. Their results demonstrate that CBR is quite an efcient tool in
It is a technique that has been proposed for a long time as a
monitoring the nancial market against possible collapse. Gener-
valid alternative to expert judgment. Searching for suitable fea-
ally, the reasoning of a CBR proceeds through the following steps
ture weights is key to the reasoning process of CBR (Auer et al.,
(Bichindaritz, 2006): (1) interpretation of a new case; (2) retrieval
2006). A strategy for specifying signicant features is to build a
of the most similar case; (3) reuse of the case to attempt to solve
search mechanism, whose algorithm can be used to determine
the problem; (4) revision of the suggested solution if necessary;
the weighted importance of various features. Genetic Algorithm
(5) retention of the solution and the new problem as a historical
(GA) is an effective search technique for solving global optimiza-
case. The measurement of distance between pairs of features plays
tion problems (Holland, 1975). Therefore, the present article aims to
an important role in CBR approaches. Essentially, more important
investigate the effect on the performance of broiler price prediction
features must be assigned larger weights than less important fea-
when GA is adopted.
tures. However, most of the CBR approaches apply equally weighted
similarity measures and do not present signicant weights for
2. Literature review
the varying features (Jorgenson, 2004; Shepperd and Cartwright,
2005).
Most of the agricultural economic research applies the econo-
GA is a useful evolutionary approach that can explore good solu-
metric model to estimate and analyze broiler price. There are
tions for complex problems. Thus, in this study, GA is adopted to
two major models, namely, time series model and simultaneous
determine the feature weights in CBR approaches for price pre-
equation model. The time series model is generally applied for fore-
diction. In addition to the equally weighted CBR, the unequally
casting or analyzing the relationship between related economic
weighted and linearly weighted mechanisms are also constructed
variables and time series data. For example, Holt and Aradhyula
to examine their potential feature weights in CBR for price predic-
(1990, 1998) have adopted the GARCH or GARM-H model to explore
tion of broilers.
the issues related to broiler wholesale price, broiler production,
and their corresponding risks. In another study, Kapombe and
3. Methodology
Colyer (1999) have analyzed the relationship between international
chicken markets and US-exported chicken price using the struc-
This study investigates and compares three weighted similar-
tural time series model. The time series model veries if the price
ity measure mechanisms for the accuracy of their broiler price
variables are non-stationary. Moreover, it tests the relationship
prediction. These mechanisms are: equally weighted CBR (ECBR),
between current and earlier prices and examines the heteroskedas-
unequally weighted CBR (UCBR), and linearly weighted CBR (LCBR).
ticity of residuals. Owing to those constraints, it is not easy to obtain
The ECBR approach applies an equal weight to distance measure for
better results. Thus, further adjustments of the variables or the
each feature in the prediction of the broiler price. Equally weighted
functional forms (e.g., rst-order difference) are required. Although
Euclidean distance is adopted to measure the dissimilarity between
the forecasting performance is acceptable, its evaluation of price
two features in the ECBR approach. The ECBR approach measures
information is still subject to the estimated time lag periods. In
the straight-line distance between two features using the values of
the simultaneous equation model, meat prices and quantities are
features normalized between 0 and 1. The overall distance between
estimated jointly. For instance, Karagiannis et al. (2000) applied
two cases of the ECBR approach is shown in Eq. (1), where n is the
the AIDS model to discuss the co-integration of major meat prices
number of features. Cp is case p to be predicted, and Ch is case h in
and their real expenditure in Greece. The various applications of
the historical case base. Cpi is the value of feature i in case p and Chi
this model were constrained to follow the theoretical models and
is the value of feature i in case h. The distance increases as similarity
their restriction conditions (e.g., Eales and Unnevehr, 1988; Golan
decreases. Furthermore, the distance metric measures the dissim-
et al., 2001). If the theoretical models are modied, their results
ilarity between two cases in terms of their features. In this study,
will not be consistent. Moreover, these econometric models are
Wi is a xed value, that is, it represents the equal weight for each
usually transformed into linear models, which are convenient for
feature i. The sum of the squares of the distance for each feature is
empirical analysis. Furthermore, some nonlinear models have been
the square of distance between two cases. Hence, the closest analog
developed in recent years, though they are still restricted by the
to a case p is the case with the minimum Euclidian distance:
functional forms. In practice, it is not easy to depict the variation 
of broiler prices through a specic functional form. As a result, the  n

proposed CBR approach attempts to overcome the limitation of the D(Cp , Ch ) =  Wi (Cpi Chi )2 (1)
econometric models and aims to provide a feasible tool for broiler
i=1
price prediction.
The uses of the CBR approaches have been evaluated and con- As shown in Fig. 1, the UCBR and LCBR are adopted when features
rmed in many price prediction elds. In 2007, Chiu and Huang have are weighted to reect the relative importance of each feature. The
applied the CBR approach for software cost prediction. Their empir- GA is adopted to probe for the suitable feature weights for dis-
ical results show that the CBR approach outperforms the widely tance measures in the CBR approach. The weights of the distance
72 M.L. Shih et al. / Computers and Electronics in Agriculture 66 (2009) 7075

Table 1
Variables of the dataset.

Feature description Maximum Minimum Mean Standard deviation

Quantity of broiler 20358947.00 11823066.00 15763360.72 2194230.04


Colorful broiler price 49.42 16.84 25.48 7.15
Price of chicks 17.87 6.00 10.94 2.69
Feedstuff price 12.04 10.31 11.08 0.48
Imported chicken quantity 10618000.00 0.00 2680847.22 2504793.22
Imported chicken price 39.32 0 23.69 5.85
Pig price 66.88 37.00 49.07 7.64
Business indicator 115.10 96.10 107.33 4.78
Consumer price index 105.47 98.51 100.65 1.64
Wholesale price index 113.74 97.33 104.03 4.70
Broiler price (dependent variable) 26.38 18.19 21.40 1.60

measures of each feature are treated as inputs, and the tness func- quantity. Other variables include related substitution products and
tion is dened as a minimization of the mean absolute error (MAE). their prices such as the price of colorful broilers, prices of pigs, and
The weight of distance measure for each feature is encoded in genes. imported chicken quantity and price. The industrial external fac-
Each chromosome is composed of a set of genes, and the collec- tors, which are closely related to macroeconomic situations, include
tions of chromosomes make up a population. Furthermore, GA can variables such as business indicator, consumer price index, and
be thought of as an evolutionary process, where a population of wholesale price index.
weights evolves over a sequence of generations. The prices of chicks and feedstuff are the major cost compo-
During each generation, the tness of each weight is calculated. nents of broiler production. If these prices change, the cost of broiler
In addition, the weights are selected for reproduction on the basis would also change and the broiler price would be subsequently
of their tness. The survival probability of a weight is proportional affected. The broiler price would drop as the supply quantity of
to its tness values. Thereafter, the reproduced weights undergo the broiler increases, hence, a balance between the demand and
recombination, which consists of crossover and mutation. Subse- supply could be reached. Given that colorful broilers and pigs are
quently, these iterative processes optimize the weights of distance important substitute products for broiler, consumers might opt for
measures for each feature. chicken if the prices of these substitute products increase. Conse-
The UCBR and LCBR approaches are based on the distance metric quently, this would induce an increase in the broiler price. It must
of Eq. (1). The GA is applied to the CBR approach to investigate a suit- be noted that the quantity of imported chicken may affect the total
able weight, Wi, for each feature. In the case of the UCBR approach, chicken supply quantity in the domestic market. For example, if the
each feature has a different weight that is obtained with respec- quantity of imported chicken increases, the broiler price would be
tive distance measure. On the basis of Eq. (1), a weighted equation forced to decrease. The price level of imported chicken could inu-
of the UCBR approach is shown in Eq. (2), where vi is a variable ence the price competitiveness of the domestic broiler. Therefore,
weight of feature i. The value of variable vi depends on the GAs the broiler price would be forced to drop if the price of imported
search results after several learning cycles. Consequently, the UCBR chicken decreases.
approach applies a corresponding constant weight for each feature: Considering the macroeconomic situations, business conditions
could reect the social consumption capacity. This in turn would
Wi = vi (2)
affect the broiler price. Suppose the business condition is ne, the
In the case of the LCBR, each feature has a different linear equation social consumption would correspondingly be encouraged. There-
to determine the feature weight that is obtained from the corre- fore, the demand for chicken might increase. As a result, the broiler
sponding distance measure. The measure varies depending on the price may possibly increase. The consumer price index is related
feature value Cpi of case p. Eq. (3) shows a linear equation of the LCBR to the purchasing power of consumers. It implies that the pur-
approach, where xi is the coefcient and vi is the constant for the chasing power of consumers would decrease when the consumer
feature i. The values of xi and vi are conrmed using GA after several price index increases. A decrease in the demand for chicken by the
learning cycles. The computed outcome of the linear equation, Wi , consumers may ensue given this situation. Moreover, this could
is the feature weight of distance measure for feature i. Therefore, translate to an indirect impact on the broiler price. To a certain
a separate linear equation for the weight of distance measure for degree, the wholesale price index would express the changes in
each feature is utilized in LCBR: farm prices, and this might be more or less associated with the
broiler price.
Wi = xi Cpi + vi (3)
What are the practical inuences of all the above-mentioned
Similarly, after the historical cases have been extracted, the next variables on the broiler price? The above question still needs to
step involves a decision on how to generate the price prediction be investigated. In other words, these factors might present some
for the case being predicted. In the context of case adaptation, the positive or negative inuences in theoretical inference. However,
authors use the mean methods. The term mean represents the their relationship might or might not be signicant in prac-
average price of the k nearest historical cases. Likewise, it is a typi- tice.
cal measure of the central tendency among the k nearest historical This study has collected 72 records from January 2000 to Decem-
cases. ber 2005 in Taiwan. As shown in Table 1, ten independent variables
are adopted for predicting broiler price in this research. The price
4. Description of dataset variables include the monthly average price, while the quantity
variables involve the aggregation of the daily transaction quantity
From the viewpoint of industrial economics, the variables that per month for every variable. The said data were collected mainly
affect the broiler price could be classied as industrial internal from government websites as well as from the Poultry Association.
and external factors. The variables in the industrial internal fac- In addition, the broiler price serves as the dependent variable. All
tors could include the price of chicks, feedstuff price, and broiler of these variables are of a continuous type. The mean value of the
M.L. Shih et al. / Computers and Electronics in Agriculture 66 (2009) 7075 73

Table 2 Table 3
The GA parameter. Results of different CBR approaches.

Parameter Value Approach MAE

Population size 10 V Training Test


Crossover rate 0.6
ECBR-1 (1 nearest neighbor) 0.868 2.589
Mutation rate 0.06
ECBR-2 (2 nearest neighbors) 0.784 2.778
Crossover method Uniform
ECBR-3 (3 nearest neighbors) 0.767 2.643
Selection method Roulette wheel
UCBR-1 (1 nearest neighbor) 0.770 1.015
UCBR-2 (2 nearest neighbors) 0.682 0.896
UCBR-3 (3 nearest neighbors) 0.659 0.912
dependent variable is 21.4 NT dollars; the minimum price is 18.19 LCBR-1 (1 nearest neighbor) 0.756 1.015
NT dollars and the maximum price is 26.38 NT dollars. LCBR-2 (2 nearest neighbors) 0.675 0.871
For this small data set, the normal practice in comparing the LCBR-3 (3 nearest neighbors) 0.626 0.836
BPN 0.585 1.670
performances of several techniques involves the use of n-folds
CART 0.503 1.280
cross-validation. It does so by creating n different training and test LR 0.589 1.176
sets so that the results are not biased. However, this data set is of a
time series mode and does not appear appropriate in terms of using
n-folds cross-validation approach. That is, each case is in conjunc- trials or when the best result did not change over the last 100 V
tion with the broiler price for a specic month between January trials.
2000 and December 2005. For example, the case that uses the pre- One of the important issues for CBR is to decide on the number
diction model trained by new cases (e.g., the data from December of similar cases considered for broiler price prediction. This deci-
2002 to December 2005) to estimate the old case (e.g., the data sion is made on a case-to-case basis and it would be reasonable to
from January 2000 to November 2002) does not make much sense. consider just a small number of cases when only small data sets are
Therefore, the whole data set is divided into pairs of training and used. In the current CBR approach, we simulated the previous work
test sets. Among the 72 records, 48 records from January 2000 to that applied a small number of three nearest neighbors for case
December 2003 are applied for training. The other 24 records from adaptation (Chiu and Huang, 2007). Table 3 shows the MAE at the
January 2004 to December 2005 are adopted for testing. That is, training and test stages of BPN, LR, CART, and three CBR approaches,
two-thirds of the whole data set is used for training and one-third based on the different numbers of case adaptations. In the different
is used for testing. numbers of case adaptations, the ECBR approach, together with the
case adaptation of the three nearest neighbors, reveals the smallest
5. The experiments and results MAE of 0.767 compared to the case adaptation of one or two nearest
neighbors. The UCBR and LCBR approaches also show the smallest
This study compared the predictive abilities of three widely used MAE of 0.659 and 0.626, respectively, with the case adaptation of
prediction models, namely, the back propagation neural network three nearest neighbors.
(BPN: Singh and Singh, 2005), the linear regression (LR: Shepperd The MAEs at training and at test stages are shown in Fig. 2.
and Cartwright, 2005), and the classication and regression trees Among the different CBR approaches, the application of GA to
(CART: Deconinck et al., 2005). The evaluation of ECBR, UCBR, LCBR, explore the suitable weight approaches (i.e., UCBR-1, UCBR-2,
BPN, LR, and CART approaches were based on the three test sets. UCBR-3, LCBR-1, LCBR-2, and LCBR-3) outperforms the equally
Table 2 summarizes the GA parameter settings in this study. The weighted CBR approaches (i.e., ECBR-1, ECBR-2, and ECBR-3) at the
results indicate that a larger population size results in a more prob- test stage. These weighted CBR approaches are also slightly better
able ability to nd a good solution. As GA is computationally very than the LR, BPN, and CART approaches at the test stage. The predic-
intensive, a trade-off must be made between the population size tion results show insignicant differences at both the training and
and execution performance. In general, the minimum effective pop- test stages for the six weighted CBR approaches (i.e., UCBR-1, UCBR-
ulation size grows in conjunction with the problem size. The GA 2, UCBR-3, LCBR-1, LCBR-2, and LCBR-3). These results show that the
parameters including population size are decided based on the pre- weighted CBR approaches are superior to the equally weighted CBR
vious study (Chiu and Huang, 2007) and on a certain amount of the approach, with respect to the data adopted in this study. It is also
authors experimentation. The GA parameters chosen for UCBR and comparable to the results obtained from the widely used prediction
NCBR approaches were 10 V organisms in the population. In this approaches of BPN, CART, and LR.
study, V is the number of variables (problem size) to be explored For the three CBR approaches, the best training model for each
for GA. The entire learning process was terminated after 1000 V CBR approach with the case adaptation of three nearest neighbors

Fig. 2. The MAE at training and test stages.


74 M.L. Shih et al. / Computers and Electronics in Agriculture 66 (2009) 7075

Table 4 Table 5
Improvements of different CBR approaches. Feature weights of independent variables.

Approach MAE Independent variable Approach

Training Test ECBR UCBR LCBR coefcient LCBR constant

Equal weight vs. unequal weight Quantity of broiler 0.100 0.000 0.000 0.00
ECBR-3 0.767 2.643 Price of colorful broiler 0.100 0.462 0.219 9.60
UCBR-3 0.659 0.912 Price of chicks 0.100 0.538 0.781 3.59
Improvement (%) 14.1 65.5 Price of feedstuff 0.100 0.000 0.000 0.00
Quantity of imported chicken 0.100 0.000 0.000 0.00
Equal weight vs. linear weight
Price of imported chicken 0.100 0.000 0.000 0.00
ECBR-3 0.767 2.643
Price of pig 0.100 0.000 0.000 0.00
LCBR-3 0.626 0.836
Business indicator 0.100 0.000 0.000 0.00
Improvement (%) 18.4 68.4
Consumer price index 0.100 0.000 0.000 0.00
Unequal weight vs. linear weight Wholesale price index 0.100 0.000 0.000 0.00
UCBR-3 0.659 0.912
LCBR-3 0.626 0.836
Improvement (%) 5.0 8.3 dent variables that are able to obtain a potentially better prediction
performance than those of other approaches.
The ndings reveal that the collaboration of the two parties,
is adopted for the evaluation of their respective prediction perfor- specically the breeding and broiler growers industries, must
mance. Table 4 shows the improvement of differently weighted involve transparency and mutual trust in order to alleviate the
CBR approaches among the training and test stages. Considering pricing argument. Furthermore, the strategic alliance between the
the improvement of unequally weighted CBR with respect to the breeding and broiler growers industries can also help maintain
equally weighted CBR, the UCBR-3 approach produces an improve- price stability and market competitive strength against the col-
ment by 14.1% at the training stage and 65.5% at the test stage. orful broiler industries. Moreover, the Poultry Association could
Considering the improvement of linearly weighted CBR with obtain a more reasonable quotation for processing plants by using
respect to the equally weighted CBR, the LCBR-3 approach dis- the prediction model. If this is successfully achieved, the exist-
plays improvement by 18.4% at the training stage and 68.4% at ing price controversy issue can be resolved. Although theoretically,
the test stage. Likewise, considering the improvement of the lin- imported chickens might inuence the market structure and impact
early weighted CBR with respect to the unequally weighted CBR, the broiler prices, this research found that this factor does not sig-
the LCBR-3 approach displays improvement by 5.0% at the train- nicantly affect the broiler price. Several reasons may be able to
ing stage and 8.3% at the test stage. Generally, each differently explain this. First, the market liberalization trend has compelled
weighted CBR approach (i.e., UCBR-3 and LCBR-3) provides signif- the local broiler industry to reduce the production quantity when
icant improvement based on the equally weighted CBR approach Taiwan joined the WTO in 2002. Second, the imported products are
(i.e., ECBR-3) at the test stage. not immediately distributed to the market place; instead they are
The improvement of the linearly weighted CBR approach with stored in frozen repositories and held for distribution according
respect to the unequally weighted CBR reveals insignicant dif- to market demands. Third, the foreign broilers are not period-
ferences at both the training and test stages. This shows that the ically imported; therefore their market impact is not salient as
unequally weighted CBR and linearly weighted CBR present similar expected.
abilities with regard to broiler price prediction with the data sets
used in this research. 6. Discussion and conclusions
A paired-samples t test is applied to investigate if the weighted
CBR outperforms the equally weighted CBR. The null hypothesis, The proposed CBR approach has demonstrated acceptable per-
which is the mean difference of two samples at the test stage, is formance when compared with most commonly used data mining
less than zero or equal to zero. In addition, the alternative hypoth- approaches, such as linear regression, CART, and neural nets. In
esis, which is the mean difference of two samples at the test stage, terms of improvement in the CBR approach, this article has exper-
is larger than zero. The signicant values at the test stage are 0.000 imented on the weight adjustment mechanism by using unequal
and 0.00 for the paired samples of ECBR versus UCBR and ECBR ver- and linear weights. However, other experiments with a nonlin-
sus LCBR, respectively. Moreover, the weighted CBR of UCBR and ear approach and similarity functions that use different search
LCBR outperform the equally weighted CBR of ECBR at a signicant techniques present some potential topics for future research. It is
level 0.05. For the paired samples of UCBR versus LCBR, the signi- generally known that pigs and colorful broilers are substitute prod-
cant value is 0.046. This is also smaller than the signicant level of ucts for broiler. However, our research ndings indicate that the
0.05. These results show that the weighted CBRs of UCBR and LCBR price of colorful broilers, more than the price of pigs, poses a greater
outperform the equally weighted CBR of ECBR. In addition, the lin- inuence on the regular broiler price. This can be explained by the
early weighted CBR of LCBR is superior to the unequally weighted various diseases that hit the pig industry in recent years, hence scar-
CBR of UCBR based on the paired-samples t test. ing off potential customers. Consequently, this leaves the price of
Table 5 shows the feature weights of independent variables for colorful broilers as the major factor that affects the pricing of the
the tree CBR approaches of ECBR, UCBR, and LCBR. All of the fea- regular broiler.
ture weights of the equally weighted CBR approach (i.e., ECBR) are The research results indicate that the proposed CBR price pre-
0.1. In reference to the unequally weighted CBR approach of UCBR, diction model can support the agriculture authority in forecasting
the feature weights are 0.538 and 0.462 for the price of chicks the broiler prices more effectively. Furthermore, the agriculture
and price of colorful broilers, respectively. The other eight feature authority can more effectively realize its role in resource alloca-
weights are all 0 for the UCBR approach. Thus, they are all insignif- tion and planning given several government support mechanisms
icant in identifying the broiler price. The linearly weighted CBR for the broiler industries. As more effective prediction results rely
approach also indicates the same situation wherein the broiler price on a complete historical data set, it is suggested that the Taiwan
is mostly inuenced by the prices of colorful broilers and chicks. agriculture authority institute a constant data collection and stor-
The proposed weighted CBR approach only requires few indepen- age mechanism for prediction and analysis. It is also believed that
M.L. Shih et al. / Computers and Electronics in Agriculture 66 (2009) 7075 75

the current research results can be improved in the future with the Holland, J. (Ed.), 1975. Adaptation in Natural and Articial Systems. University of
collection of more data. Michigan Press, Ann Arbor, MI, USA.
Holt, M.T., Aradhyula, S.V., 1990. Price risk in supply equation: an application of
GARCH time-series models to the U.S. broiler market. Southern Econ. J. 57 (1),
References 230242.
Holt, M.T., Aradhyula, S.V., 1998. Endogenous risk in rational-expectations commod-
Auer, M., Trendowicz, A., Graser, B., Haunschmid, E., Bif, S., 2006. Optimal project ity models: a multivariate generalized ARCH-M approach. J. Empirical Financ. 5,
feature weights in analogy-based cost estimation: improvement and limitations. 99129.
IEEE Trans. Softw. Eng. 32 (2), 8392. Jorgenson, M., 2004. A review of studies on expert estimation of software develop-
Bichindaritz, I., 2006. Case-based reasoning in the health sciences. Artif. Intell. Med. ment effort. J. Syst. Softw. 70, 3760.
36, 121125. Karagiannis, G., Katranidis, S., Velentzas, K., 2000. An error correction almost ideal
Chiu, N.H., Huang, S.J., 2007. The adjusted analogy-based software effort estimation demand system for meat in Greece. Agric. Econ. 22, 2935.
based on similarity distances. J. Syst. Softw. 80, 628640. Kapombe, C.M., Colyer, D., 1999. A structural time series analysis of US broiler
Council of Agriculture, 2005. Agriculture Statistics Yearbook. Executive Yuan. Council exports. Am. J. Agric. Econ. 21, 295307.
of Agriculture, Taipei, ROC. Oh, K.J., Kim, T.Y., 2007. Financial market monitoring by case-based reasoning. Expert
Deconinck, E., Hancock, T., Coomans, D., Massart, D.L., Heyden, Y.V., 2005. Classica- Syst. Appl. 32, 789800.
tion of drugs in absorption classes using the classication and regression trees Shepperd, M., Cartwright, M., 2005. A replication of the use of regression towards
(CART) methodology. J. Pharm. Biomed. Anal. 39, 91103. the mean (R2M) as an adjustment to effort estimation models. In: Proceedings
Eales, J.S., Unnevehr, L.J., 1988. Demand for beef and chicken products: separability of the 11th IEEE International Software Metrics Symposium (METRICS 2005), pp.
and structural change. Am. J. Agric. Econ. 70 (3), 521532. 38138.
Golan, A., Perloff, J.M., Shen, E.Z., 2001. Estimating a demand system with non- Singh, S., Singh, M.P., 2005. Prediction of missed number from pin code of ATM card
negativity constraints: Mexican Meat Demand. Rev. Agric. Econ. 83 (3), 541 with back propagation method of articial neural network. Appl. Math. Comput.
550. 171, 152162.

Вам также может понравиться