

doi:10.1016/j.eswa.2005.09.086


Neural networks for animal science applications: Two case studies

C. Fernández
a
, E. Soria

b,*, J.D. Martı́n
b
, A.J. Serrano

b

a
Departmento. Producción Animal, Facultad de Ciencias Experimentales y de la Salud, Universidad Cardenal Herrera CEU,

Edificio Seminario s/n. 46113 Moncada, Valencia, Spain
b

Digital Signal Processing Group, Department Enginyeria Electrònica, Escola Tècnica Superior d’Enginyeria,

Universitat de València, C/ Dr. Moliner 50, 46100. Burjassot, Valencia, Spain

Abstract

Artificial neural networks have shown to be a powerful tool for system modelling in a wide range of applications. In this paper, we focus on

neural network applications to intelligent data analysis in the field of animal science. Two classical applications of neural networks are proposed:

time series prediction and clustering. The first task is related to the prediction of weekly milk production in goat flocks, which includes a

knowledge discovery stage in order to analyse the relative relevance of the different variables. The second task is the clustering of goat flocks; it is

used to analyse different livestock surveys by using self-organizing maps and the adaptive resonance theory, thus obtaining a qualitative

knowledge from these surveys. Achieved results show the usefulness of neural networks in two animal science applications.

q 2005 Elsevier Ltd. All rights reserved.

Keywords: Neural networks; Multilayer perceptron; Self-organizing map; Animal science
1. Introduction

Artificial neural networks (ANNs) have been applied to a

huge amount of fields during last years (Arbib, 2003; Hu,

2003). However, there are some fields in which the use of

ANNs is still scarce; animal science is one of these fields. This

lack of ANN applications in animal science is quite

paradoxical, as data analyses are usually carried out in this

field, and ANNs have shown to be more powerful than classical

statistical methods to carry out this kind of tasks (Arbib, 2003;

Bishop, 1995; Ripley, 1996).

Two examples are used in this work in order to show the

ability of ANNs to solve animal science problems: a time series

prediction (Makridakis, 1997; Weigend, 1993) and a clustering

application (Theodoridis, 1999). The multilayer perceptron

(MLP) (Bishop, 1995) is the neural model used for time series

prediction, whereas the neural models used for clustering are

the Kononen’s self-organizing map (SOM) (Kohonen, 2000)

and a network based on the adaptive resonance theory (ART),

the ART2 network (Fausset, 1994; Jang, 1996).

The first problem tackled in this paper is related to milk

yield (MY) weekly prediction by using current animal factors
0957-4174/$ - see front matter q 2005 Elsevier Ltd. All rights reserved.

doi:10.1016/j.eswa.2005.09.086

* Corresponding author.

E-mail address: emilio.soria@uv.es (E. Soria).
(body weight, diet, litter size, number of lactation, etc.), as well

as the current MY prediction (an autoregressive model). MY

weekly prediction becomes more useful for management

purposes than the analysis of the different factors which can

affect MY at the end of lactation.

The clustering application is completely different since a

qualitative model is desired in order to extract useful

knowledge from data. Goat production in Murcia Region

(Spain) requires a deeper knowledge of their farming

characteristics and practices. Knowledge of MY prediction

is important because is the main income for farmers, but

without a good feeding program is difficult to achieve their

goals. The nature and availability of biomass of semiarid

Spanish lands may limit the animal welfare and pro-

ductivity, and thus, lower milk production may be achieved,

as well as a decrease in kid survives and income for selling

milk. Supplementation during the year with forages, by-

product, concentrates and/or total mixed ration (TMR), is a

common practice when pasture is scarce, as happens in

Murcia Region. Ninety-four flocks from Murcia Region are

used to analyse the relationship between food management

practices. The herd food management practise is obtained

by surveys to farmers.

The remainder of the paper is outlined as follows. Section 2

shows the neural methods used in this study. Results achieved

in different applications are presented in Section 3, ending up

the paper with some conclusions in Section 4.
Expert Systems with Applications 31 (2006) 444–450
www.elsevier.com/locate/eswa

http://www.elsevier.com/locate/eswa


Fig. 1. Schematic of a neuron model.

C. Fernández et al. / Expert Systems with Applications 31 (2006) 444–450 445
2. Neural models

2.1. Multilayer perceptron

A multilayer perceptron (MLP), is an ANN, formed by

elementary processing units, the so-called neurons. A typical

neuron model is shown in Fig. 1 (Haykin, 1999).

The main components of the model are:

Sum function. It carries out a linear combination of the

neuron inputs through the use of a set of coefficients, known

as synaptic weights. Being wZ[w0, w1,., wk] the vector of
coefficients, and xZ[1, x1,., xk] the input vector, the sum
function is given by the scalar product of both vectors.

When using a neural model, the goal is to find the optimal

synaptic weights to solve the problem. The process of

searching the optimal weights is known as network learning.

Activation function. It is a non-linear function which gives

the network its non-linear nature. The most used activation

functions are the sigmoid function (its values ranging

between 0 and 1) and the hyperbolic tangent, which ranges

between K1 and C1.

As it can be inferred from Fig. 1, a neuron without an

activation function is equivalent to a multi-variant analysis.

Therefore, a non-linear combination should be more powerful

than a multi-variant analysis (Ripley, 1996).

Neurons are arranged in layers to form an MLP. The

first layer is known as input layer, and the last one is

called output layer. All the other layers are called hidden

layers (Arbib, 2003). This kind of arrangement enables

the neuron outputs to be used as inputs to neurons of

following layers (non-recurrent network) and/or previous
Fig. 2. Multilayer perceptron. Dotted line
layers (recurrent networks). Fig. 2 shows a typical MLP

structure.

Several remarks should be made regarding the learning

process (Arbib, 2003; Bishop, 1995; Haykin,1999):

The learning rate must be chosen appropriately. The

learning algorithm is based on finding the closest minimum,

being this learning rate a parameter which measures the

speed of approaching the minimum.

The network architecture must also be chosen appropriately.

While the number of neurons in the input layer and the

output layer is given by the problem, the number of hidden

layers, and the quantity of neurons in each hidden layer must

be chosen depending on the particular problem.

Due to the iterative nature of the learning process, it is

necessary to choose the initial values of the synaptic

weights. Different initial values are usually tested in order to

achieve an optimal model.

In order to obtain a model with validation capabilities, data

is split into two data sets: a training set used to obtain the

neural model, and a validation set used to test the behaviour

of the model with data different from the training set used to

obtain the model.

Two different approaches can be carried out to train the

network. The first approach is based on updating the

synaptic weights for each pattern of the data set (on-line

approach) whereas the other approach is based on updating

the synaptic weights once for all the training data set (batch

approach).
2.2. Self-organizing map (SOM)

The self-organizing map is a neural network proposed by

Teuvo Kohonen in 1984 (Haykin, 1999; Kohonen, 2000).

Neurons are arranged in two layers: an input layer, formed by n

neurons (one neuron for each input variable) in Fig. 3 and an

output layer in which information is processed; this second

layer is usually arranged in a two-dimensional structure.

Neurons of the output layer are characterised by a weight

vector with the same dimension as the input vector. For

instance, neuron i,j (i-th row, j-th column) is characterised by

the weight vector wij Z½w
1
ijw

2
ij:::::::::w

n
ij�. Similar input patterns

are mapped close each other in the output layer (Kohonen,
s stand for recurrent neural systems.


Fig. 3. SOM architecture with n-dimensional inputs.

C. Fernández et al. / Expert Systems with Applications 31 (2006) 444–450446
2000). Algorithm procedure can be summarized, as follows

(Haykin, 1999; Kohonen, 2000):

Weight initialisation.

Choice of an input pattern xZ x1 x2:::::::::xn
� �

.

Measurement of the similarity between weights and inputs.

If the Euclidean distance is taken into account, then the

similarity measure is given by dðwij; xÞZ
PM

kZ1
ðwkij KxkÞ

2

The most similar neuron to the input pattern is called best

matching unit (BMU).

Synaptic weights are updated as wstðnC1ÞZwstðnÞC
a,hðBMU; wstÞ,ðxKwstÞ, where a(n) is the learning rate
and h(n) is known as neighbourhood function. The value of

this function depends on the distance between the BMU and

the neuron to be updated, the closer the two neurons the

higher the value of this function. Moreover, this function is

the responsible of preserving the topological relationships

among input patterns (Kohonen, 2000).

The previous steps are performed a predetermined number

of iterations. When this number is reached, the learning

algorithm is stopped. While the number of iterations is

lower than the predetermined value, go to step 2.

Once the map training is finished, the visualisation of the

two-dimensional map provides a qualitative information about

how the input variables are related each other for the data set

used to train the map. SOM is a visualisation tool rather than a

clustering tool, although it is possible to obtain clusters of

similar patterns from the two-dimensional map. Nevertheless,

another model is proposed in order to carry out this clustering

task, namely, the adaptive resonance theory (ART).
2.3. Adaptive resonance theory (ART)

This model is based on analysing the similarity between

input patterns and cluster prototypes. If a pattern is similar

enough to a certain cluster prototype, then the prototype is

updated in order to incorporate information from this pattern. If

the similarity is not sufficient, then a new cluster is formed,

whose prototype is given by this pattern. This model solves a

usual problem of other neural models, namely, the corruption
of the information already learnt by the contribution of new

patterns (Arbib, 2003). In particular, we used the ART2

network which allows to use this theory with continuous-

valued input vector (Fausset, 1994). Cluster prototypes have

the same dimension as input vectors. ART2 procedure can be

stated as follows (Kung, 1993):

Weight initialisation.

Given an input pattern x, the Euclidean distance between

this pattern and every neuron is computed. Therefore, the

distance between the j-th neuron and the input pattern is

given by dj Zjjwj Kxjj.
The most similar neuron to the input pattern is selected in

order to analyse whether it is a good-enough representation

of the input pattern. This neuron is denoted as v.

A vigilance test is performed in order to test whether the

prototype is a good enough representation of the pattern:

jjwv,xjjOp where r is the vigilance parameter (0!r!1).
If the vigilance test is not accomplished, then a new neuron

is created, whose weights are given by the input pattern.

If the vigilance test is accomplished, then the weight vector

of the winner neuron is updated:

wv Z
x Cwv,Nv

Nv C 1

where Nv is the number of data points in the winner node v

so far. This means that if there are many points in the

neighbourhood of the winner, then it is hard to move the

winner. The vigilance parameter controls the desired degree of

similarity between the weight vector and the input pattern; low

values involve that a low similarity is needed, and therefore, a

low number of clusters is produced. On the contrary, high

values involve a high number of prototypes.
3. Practical applications

3.1. Milk yield prediction

Milk yield (MY) is a key factor in the management of goat

livestock since it becomes the main farmer income. MY is

important to identify which goats are the best milk producers,

and also to analyse the current state of a certain livestock.

Therefore, MY prediction is very helpful as it enables the

farmer to make decisions about livestock management before

the end of lactation, when decisions are usually made.

A homogenous group of dairy goats within a commercial

farm (Excamur, S.L.) was selected in this trial. This farm is a

member of ACRIMUR (Murciano-Granadina Goat Breeder

Association—Asociación Española de Criadores de la Cabra

Murciano-Granadina). The selected farm is a representative

sample of these kinds of farms in the south-eastern Spain.

The data set used in this work was formed by 36 goats. They

were split into two data sets: a training set (26 goats) to obtain

the prediction model, and a validation set (10 goats) in order to

guarantee the generalisation capabilities of the network. Both

data sets were similar with regard to the diet followed by goats


Table 1

Mean values and standard deviations of the input variables for both training and

validation sets

Lactation

number

(*)

Number

of kids

Control

number

Metabolic

weight

(kg3/4)

Milk yield

(kg)

Training

Mean 4.94 1.82 8.13 16.91 2.20

Std 1.37 0.58 3.52 1.25 0.80

Validation

Mean 4.80 2.10 9.10 17.36 2.06

Std 1.25 0.54 3.42 1.63 0.64

* Lactation number indicates lactation period.

Table 2

Root mean-square error (RMSE) obtained by the simple models as well as by

the best neural models in MY prediction. Net(1) has an architecture 6!3!1

and online learning, Net(2) 6!3!1 and batch, Net(3) 6!3!3!1 and batch,

and finally, Net(4) 6!4!4!1 and online

Training set (kg) Validation set (kg)

Model1 0.47 0.42

AR(1) 0.44 0.39

AR(2) 0.44 0.40

Net(1) 0.37 0.31

Net(2) 0.37 0.31

Net(3) 0.37 0.31

Net(4) 0.36 0.31

C. Fernández et al. / Expert Systems with Applications 31 (2006) 444–450 447
in order to avoid biased models. A weekly follow-up was

carried out for every goat, being the study period of 22 weeks.

The most important variables that could affect MY were

taken into account (Gipson & Grossman, 1990). These

variables can be easily obtained in a flock which is subject to

a milk control process. The used variables were the following:

number of lactation, number of kids, time between parturition

and the first milk control, type of diet
1
, metabolic weight

(Klieber, 1961), and current MY. Table 1 shows the mean

values and the standard deviations values of the input variables;

similar values for both data sets show the suitability of the data-

splitting criterion, which guarantees unbiased models.

Three simple models were proposed in order to have a fair

reference for model comparison:

A trivial model (Model1) which uses the following

prediction rule: m(tC1)Zm(t), being m(k) the variable to
be predicted (MY).

Two autoregressive models: AR(1) which takes into account

the previous value of the prediction variable, and AR(2),

which considers the two previous values of this variable.

With regard to neural models, different architectures and

weight initialisations were proposed. This procedure is due to

the learning algorithm used, the Levenberg-Marquardt algor-

ithm, which is based on finding the closest minimum to the

initial weights, so that, if this is a local minimum, then a non-

optimal solution is achieved (Luenberger, 1984). This

algorithm shows a better performance than other more widely

used algorithms, such as, the classical backpropagation

algorithm (BP). With regard to hidden nodes, if only one

hidden layer was taken into account, the number of neurons

ranged between 2 and 15, whereas if the two hidden layers

were taken into account, the number of neurons in each layer

ranged between 2 and 6, in order to avoid overfitting (Haykin,

1999). Moreover, the stopping criterion was based on cross-

validation (Bishop, 1995; Haykin, 1999).

A comparison of the results achieved by the different

models is shown in Table 2.
1
In this work, two diets were used to fulfil nutritive requirements during

lactation, being the difference between these two diets the ingredients of each

one.
Fig. 4 shows the MY predicted by the neural model Net(4)

compared to the desired signal, for both training and validation

sets. An accurate prediction is achieved in a global point of

view. Nevertheless, it is also important to analyse whether the

prediction is accurate for every particular goat; therefore,

results for individual goats are also shown in Fig. 5.

Fig. 5 shows that the prediction achieved by Net(4) is also

accurate for individual goats, except for a small amount of

goats; for instance, goat #24 in the training set and goat #7 in

the validation set.

3.2. Qualitative analysis of goat farms by SOM

Section 3.1 carries out a quantitative analysis to predict MY

in goat farms. Nevertheless, qualitative analyses are also

important since they are very helpful to understand how the

models work, and thus, they can be used for knowledge

discovery. Sections 3.2 and 3.3 show the use of clustering

techniques for analysing the relevance of variables involved in

this problem as well as to identify characteristic behaviours in

the management of goat farms.

Feeding becomes the most important factor to take into

account when a new stockbreeding activity is initiated since it

typically involves more than half of farming expenses

(Haenlein, 2001). This is the reason why farmers filled in a

survey on their feeding procedures in order to study feeding

management. The usefulness of this information stems from

the fact that it allows an overall estimation of farms, thus trying

to amend mistakes for improving farm development and

competitivity.

Data used in this study are referred to 1999 and were

obtained between 2001 and 2002 from surveys returned by 94

farm owners dedicated to goat production. The surveyed farms

comprised approximately 15% of the total official census of

Murciano-Granadina breed farms in Murcia Region. Eleven

variables were studied. The first two variables were related to

the geographical location in Murcia Region (6 counties (V1):

1-Altiplano, 2-Campos de Cartagena, 3-Noroeste, 4-Rı́o Mula,

5-Valle del Guadalentı́n, 6-Vega del Segura) and herd size (V2:

number of goats). The other nine variables were related to

nutrition management and milk income. Three variables were

obtained asking farmers whether they supplement the diet for

different physiological stages (replacement feeding V3: No(0)

/Yes(1); lactation feeding V4: No(0)/Yes(1); gestation feeding


Fig. 4. MY predicted by Net(4) and desired signal for both training and validation sets.

C. Fernández et al. / Expert Systems with Applications 31 (2006) 444–450448
V5: No(0)/Yes(1)). The sixth variable was obtained by asking

farmers whether they have any nutrition support by external

personnel (V6: No(0)/Yes(1)). The variable type of feeding

(V7) has three possible answers; diet elaborated by farmers

(coded as 0), farmers who buy compound feed (coded as 1) and

farmers who buy total mixed ration (coded as 2). The variable

peak lactation feeding (V8) has three options as well; No

(coded as 0), Yes (coded as 1) and animal for meat production

(coded as 2). Body condition score (V9) is a subjective
Fig. 5. MY prediction for the individual goats which form the training set and the va

(MAE) are used as accuracy measures.
measurement that farmers tend to use in order to evaluate

whether nutrient supplementation appears to be necessary

(No(0)/Yes(1)). Milk income (V10) (V/ goat) was evaluated
for each farm, and finally the last variable taken into account

was the use of Byproduct (V11: No(0)/Yes(1)) to feed animals.

Our aim was to extract knowledge from these variables.

SOM was applied to carry out this task since it preserves

topological relationships among data while mapping data into a

two-dimensional map. An SOM formed by 8!6 neurons with
lidation set. The root-mean-square error (RMSE) and the Mean Absolute Error


Fig. 6. Grey-scale representation of the components of the vectors which

represent each neuron in the map. The darker the colour the lower the values of

the components.

C. Fernández et al. / Expert Systems with Applications 31 (2006) 444–450 449
hexagonal architecture (each neuron has six neighbours) was

used. Data was normalised (mean zero and variance unity) in

order to avoid unbiased models due to the different range of

values of the variables. The batch mode was used to train the

network and a Gaussian neighbourhood function was chosen.

Fig. 6 shows the components of the vectors which represent

each neuron in the map. There are several zones of the map that

should be highlighted. For instance, those variables in which

both ‘replacement feeding’ and ‘lactation feeding’ appear to be

relevant, are mapped into the lower parts of the map in light

colour. These zones of the maps also correspond, within

variable ‘type of feeding’, to those farms using compound feed

(grey colour and intermediate part of the map) and total mixed

ration (light colour and lower part of the map). On the contrary,

the dark area in the upper part of the map represents farmers

who make their own diet. To sum up, the lower part of the map

with light colour represents farmers which consider groups of

goats within their herds, and these groups are fed according to

the physiological stage (mainly replacement and lactation).

Moreover, these farmers tend to buy to food companies. The

light area is also present in the variable ‘nutrition support’, i.e.,

farmers who buy to food companies, are also supported about

their technical decisions; moreover, it can also be appreciated

that ‘milk income’ also shows light colours, which indicates
Table 3

ART2 clustering. Cluster prototypes and percentage of patterns assigned to each cl

Cluster V1 V2 V3 V4 V5 V6

1 2.12 185.85 0.97 0.79 0.33 0.52

2 3.43 124.07 0.13 0 0.13 0.33

3 4.85 251.58 1 1 0.81 0.69

4 5 53 0.33 0 0.33 0.33

5 1 1450 1 1 1 0

6 4 800 1 1 0 0
more income from milk yield. The use of feeding sub-products

(variable ‘byproducts’) is biased to the right map of the map,

including both farmers who buy to food companies and those

who make their own rations. The analysis of variable ‘gestation

feeding’ shows that this topic is still in an early development,

even though its physiological relevance. In spite of farmers do

carry out an individualised feeding of animals which are being

milked, they do not take into account individual factors such as

the four weeks of the ‘peak lactation feeding’ (maximum milk

yield). Moreover, only a few farmers take into account the

‘body condition score’.
3.3. ART clustering of goat farms

The main advantage of ART over other clustering

algorithms stems from the fact that it does not need to know

the number of clusters in advance, but it can find the natural

number of clusters once chosen the vigilance parameter r

(Theodoridis, 1999). ART2 models were trained, finding an

optimal value of rZ0.98. Six clusters were found by the
algorithm, whose prototypes are shown in Table 3.

Table 3 shows that three clusters include almost all the

patterns, namely, clusters #1, 2 and #4. Cluster #1 include 35%

of the patterns which belong to the county Campos de

Cartagena. Farms of this cluster show a medium size and

also medium income from MY (308 V per goat and lactation).
Some farmers produce their own feeding whereas others buy to

food companies, thus taking advantage of the technical support

provided by companies. They carry out a management by

groups of goats, depending on their MY capabilities.

Cluster #2 include 32% of the patterns which belong to the

north-western area. This cluster includes smaller farms and

their income is also lower than Cluster #1. These farms do not

receive technical support, but they produce their own food and

their management is the same for all the goats. Therefore, this

cluster shows a management quite poorer than that of cluster

#1, thus suggesting that an improvement should be taken into

account for a considerable percentage of farms in Murcia

Region.

Cluster #4 is located in the county Valle del Guadalentı́n

and it includes small farms. They do not show any income from

milk sale because these farms are focused on meat instead of

milk. Their management does not consider differences among

goats, and therefore, this cluster represents the poorest

management among farms of Murcia Region.
uster

V7 V8 V9 V10 V11 % Data

1.48 0.12 0.48 308.48 0.67 35.11

1.27 0.33 0.47 141.8 0.3 31.92

2.12 1 0.27 267.27 0.35 1.06

1.33 2 0 0 0.33 27.66

3 0 0 0 1 3.19

3 1 1 355 1 1.06


C. Fernández et al. / Expert Systems with Applications 31 (2006) 444–450450
The other clusters are not representative enough since they

involve a very small percentage of data.
4. Conclusions

Neural networks have been used in two animal science

applications. Results have shown their suitability to be used in

this field, in which the quality and accuracy of the models is

essential to increase farm returns.

With regard to MLP, achieved results have been better than

those obtained with classical easy models in terms of accuracy

(prediction improvement: 0.11 kg/goat/lactation). This

involves a prediction improvement of 22 kg/lactation for a

medium size herd formed by 200 goats, which may improve

herd management. Moreover, neural networks have provided

an estimation of farm locations, which could be used to avoid

mistakes and improve their development and competitivity.

With regard to SOM, maps have shown that farmers tend to

carry out a management based on groups of goats, which has

shown that feeding management of farms focused on dairy

produce is more adequate than that of farms focused on meat

production. This adequate management has also included

technical support and it has involved an income increase from

MY. These conclusions are similar to those obtained in

previous works, which studied goat herds from other Spanish

regions, such as Andalusia (Castel, Mena, Delgado-Pertiñez,

Carmuñez, Basalto & Caravaca et al., 2003; Delgado-Pertiñez,

Alcalde, Guzmán-Guerrero, Castel, Mena & Caravaca, 2003)

or Catalonia (Milán, Arnalte, & Caja, 2003).

With regard to ART, it has produced three relevant

clusters which showed three kinds of herd management’s

practices in Murcia Region. Basically, these clusters

represented farms which were characterised by a policy

based on dairy goats with supplementary feeding, dairy goats

without supplementary feeding, and finally, meat production.

The relevance of supplementary feeding has turned out to be

very importance since profits were raised (between 140 V
and 310 V per goat)
References

Arbib, M. (2003). The handbook of brain theory and neural networks.

Cambridge, MA: MIT Press.

Bishop, C. M. (1995). Neural networks for pattern recognition. Oxford:

Clarendon Press.

Castel, J. M., Mena, Y., Delgado-Pertiñez, M., Carmuñez, J., Basalto, J.,

Caravaca, F., et al. (2003). Charactarization of semi-extensive goat

production systems in southern spain. Small Ruminant Research, 47,

133–143.

Delgado-Pertiñez, M., Alcalde, M. J., Guzmán-Guerrero, J. L., Castel, J. M.,

Mena, Y., & Caravaca, F. (2003). Effect of hygiene-sanitary management

on goat milk quality in semi-extensive systems in spain. Small Ruminant

Research, 47, 51–61.

Fausett, L. V. (1994). Fundamentals of neural networks. Englewood Cliffs, NJ:

Prentice-Hall.

Gipson, T. A., & Grossman, M. (1990). Lactation curves in dairy goats: A

review. Small Ruminant Research, 3, 383.

Haenlein, G. F. W. (2001). Past, present and future perspectives of small

ruminant dairy research. Journal of Dairy Science, 84, 2097–2115.

Haykin, S. (1999). Neural networks: A comprehensive foundation. Englewood

Cliffs, NJ: Prentice Hall.

Hu, Y. H., Hwang, J. N., & Hwang, J. N. (2003). Handbook of neural network

signal processing. London: Academic Press.

Jang, J. S., Sun, C. T., & Mizutani, E. (1996). Neuro-fuzzy and soft computing:

A computational approach to learning and machine intelligence. Engle-

wood Cliffs, NJ: Prentice Hall.

Klieber, M. (1961). The fire of life (an introduction to animal energetics). New

York: Wiley.

Kohonen (2000). Self-organizing maps. New York: Springer.

Kung, S. Y. (1993). Digital neural networks. Englewood Cliffs, NJ: Prentice

Hall.

Luenberger, D. G. (1984). Linear and nonlinear programming. Reading, MA:

Addison-Wesley.

Makridakis, S. G., Wheelwright, S. C., & Hyndman, R. J. (1997). Forecasting:

Methods and applications. London: Wiley.

Milán, M. J., Arnalte, A., & Caja, G. (2003). Economic profitability and

typology of Ripollesa breed sheep farms in Spain. Small Ruminant

Research, 49, 97–105.

Ripley, B. D. (1996). Pattern necognition and neural networks. Cambridge:

Cambridge University Press.

Theodoridis, S., & Koutroumbas, K. (1999). Pattern recognition. New York:

Academic Press.

Weigend, A. S., & Gershenfeld, N. A. (1993). Time series prediction:

Forecasting the future and understanding. Proceedings of the NATO

advanced research workshop on comparative time series. Reading, MA:

Addison Wesley.


	Neural networks for animal science applications: Two case studies
	Introduction
	Neural models
	Multilayer perceptron
	Self-organizing map (SOM)
	Adaptive resonance theory (ART)

	Practical applications
	Milk yield prediction
	Qualitative analysis of goat farms by SOM
	ART clustering of goat farms

	Conclusions
	References