ANN is nonlinear model that is easy to use and understand compared to statistical methods. ANN is non-paramateric model while most of statistical methods are parametric model that need higher background of statistic. ANN with Back propagation (BP) learning algorithm is widely used in solving various classification and forecasting problems. Even though BP convergence is slow but it is guranteed. However, ANN is black box learning approach, cannot inteprete relationship between input and output and cannot deal with uncertainties. To overcome this several approches have been combined with ANN such as feature selection and etc.

Meanwhile Fuzzy is quite good in handling uncertainties and can inteprete relationship between i/o by producing rules. Therefore, to increase the capability of Fuzzy and ANN, hybridization of ANN and fuzzy is usually implemented.

Mathematical models are typically used to model a system when the system is not so complicated, but when the complexity of a system is increased other methods should be used for modeling. Fuzzy theory is typically used when the behavior of system is the most complicated or the linguistic rules are needed to define the behavior of a system. But in a condition between above-mentioned ANN are good modeling method which can produce good results.

Artificial neural networks (ANNs) have some interesting properties that made these family of machine learning algorithms very appealing when confronting difficult patter-discovery tasks. Why use ANNs instead of, say, fuzzy engines? It really depends on the problem. Sometimes ANNs will perform better than other techniques. Some others, ANNs will perform worse. The trick is to know the domain of competence (or sweet spot) for every family of algorithms. However, this is a difficult task so we keep trying with different approaches.

The models you mentioned can be used to describe relatively simple problems. ANNs are used to treat super complicated problems, in which too many variables to be simplified in a model.

Generally, The major advantage of using artificial neural networks is ....

a) Training large amount of data sets ...

b) The output performance will depend upon the trained parameters and the data set relevant to the training...

Apart from the previously mentioned advantages of the ANNs algorithms, they have the advantage to be easily implemented in parallel architectures (i.e. in multicore processors or systems with GPUs). That reduces drastically the processing time compared to other kind of algorithms, achieving similar results.

They are specially interesting the Deep Neural Networks, a kind of ANNs with a huge number of units and layers, that achieves high level of learning with low supervision.

Artificial neural network is a powerful data-driven, self-adaptive, flexible computational tool having the capability of capturing nonlinear and complex underlying characteristics of any physical process (e.g.damage detection) with a high degree of accuracy. However, they are generally regarded to behave as “black-box” systems. Due to this, the user cannot explain how learning from input data was

performed.

The main advantages of using Artificial Neural Networks (ANN) include: it can handle large amount of data sets; it has the ability to implicitly detect complex nonlinear relationships between dependent and independent variables; it has ability to detect all possible interactions between predictor variables; etc.

Since the pros and cons of ANN are already described by the colleagues above, I only want tp point out, that ANN should not be used if there is a chance to described a model mathematically. Modern optimization algorithm, i.e. INTELLIGENT search methods are more powerfull.

I see that ANN may be looking appealingly. To my experiences which however lie many years ago were that paper presenting results obtained by ANN were generally more promising than serious. Because of getting funds?

My own experiences, when I was attracted by this "convincing" idea to use this "all-problem solver" ANN, were mostly negative or at least not better than nonlinear

least square methods. As I have said, I gave up to pursue using ANNs, for example,

to approximate near optimal feedback (closed loop) controls from bundles of optimal open loop trajectories (e.g. in differential games). Intelligent mathematical methods were generally much better.

However I should not conceal my only positive example, to solve a problem, where patients had to be guided through the different stations of a hospital, so that the stations are used to full capacity as much as possible, while the waiting times of patients are reduced as much as possible. Many of the rules of this model could only be modelled in a weak non-mathematical language.

Hi,

if you have interest in how this algorithm works, please check this youtube video

http://youtu.be/LCzufhtIFnY

Regards

I am using nonparametric regression such as Nadaraya-Watson estimator instead multilayer perceptron. Try it. It is very simple and clear. This is nonlinear model with small number of parameters to estimate. And you can interpret these parameters. You can control the variance-bias tradeoff by the parameters.

I fully agree with the earlier ans.

1. First check really ANN is required to model the given problem.

2. The advantage area: ANN Does not require any model apriori. The model need not assume any model structure before starting the ANN model.

3. It can be used for non-linear problems?

4. It is a non parametric method, thus eliminates the error in parameter estimation.

In-spite of its larger advantage:

Disadvantage:

1. It cannot extrapolate the results!

2. Extracting the knowledge (weights in ANN) is very difficult (few papers are available and they have classified ANN as grey box model rather than a black-box model)

I too agree with earlier ans.

If indeed you cannot find a closed formulation of your problem and want to insure better generalization capabilities than what ANN offer, i suggest you try kernel based Support Vector Machines.

Dear Sir,

It depends on the kind of the problem you want to solve. You have some others technics for classification or clustering (supervised and unsupervised).

All sounds like good advice, ANNs are capable of modeling highly nonlinear systems, they are black boxes, work well for large data sets. I use ANNs as a very convenient method to approximate just about any relationship between predictors and predictands provided that there is sufficient data and it is a nonlinear problem. That said caution should be used e.g. not using ANNs for extrapolation as mentioned above and watch out for potential overfitting. Discipline with the handling/splitting of the data for parameters estimation and model performance evaluation is very important and be ready to drop this approach if there is not enough data.

One rule that has served me well with ANN is "justify complexity", e.g. whether to use or not a Neural Net and when selecting the topology of the ANN. This includes comparing with multilinear regressions when appropriate or other mathematical models. Comparing with SVM, RF or other techniques sounds good. I Typically use ANN and/or RF. From readings and experience, if one is mostly interested in performance, ANNs are regularly among the most performing although the experience of the modeler with a particular technique is likely very important.

On the topic of black box, a Neural Net is typically (or at least should be) used for the modeling of non linear problems. For a substantially non linear problem I don't think one can attribute physical meaning to most model parameters, except possibly locally. So part of the black box nature of the neural net is associated to the non linear problems rather than being intrinsic to ANNs. I had discussions with colleagues using AI on this topic, I am wondering if others more mathematically inclined on this post have further insights to share?

ANN should be used with care. Only some problems are good candidate for ANN, such as learning balance on a bicycle as humans do, or where the input set is limited and well defined. It is clear that ANN for example can not find discontinuity if the data set does not have data near it. So in general I do not agree that ANN can be described as a tool for intelligence, since dealing with traditional even nonlinear mapping is not intelligence. Intelligence as human perceive carry the burden of generating a hypothesis to understand data. Some people can speak language from just living with people speaking it, but surely they will fail to speak the language in formal meetings unless they figure out the linguistic rules.

Dear Khairul,

Concerning advantages and disadvantages of using artificial neural networks, please see:

http://en.wikipedia.org/wiki/Artificial_neural_network

http://www.ask.com/question/advantages-of-artificial-neural-networks

http://www.ncbi.nlm.nih.gov/pubmed/8892489

http://www.learnartificialneuralnetworks.com/introduction-to-neural-networks.html

Best!

Yaakov

Neural networks often exhibit patterns similar to those exhibited by humans. However this is more of interest in cognitive sciences than for practical examples.

Neural networks are quite simple to implement (you do not need a good linear algebra solver as for examples for SVNs).

The colleague described ANN being black box model as a disadvantage, which I see an advantage coz the user does not have to be the expert of ANN to use it. One can apply ANNs to the problem at hand that does not have any clear mathematical relationship but expected to have one without going into ANN's intricacies.

ANN is easily accessible since it is directly available in the MATLAB toolbox. Also you don't need to learn any coding to use the toolbox. In terms of performance, it is clearly not the best.

ANN Matlab tool is simply to use and works well enough at the first analysis approach. I think that simplicity is a precious SW quality

We have used Artificial Neural Networks at the work indicated in the link below:

https://www.researchgate.net/publication/3840338_Security_management_against_cloning_mobile_phones?ev=prf_pub

ABSTRACT: This work presents the development of a distributed security

management system for telecommunication networks. The system consists in

reducing the use of cloned mobile telephones using three main

techniques: (1) an ISO formal technique (LOTOS) is used to specify and

validate the system; (2) a pattern recognition technique is used to

classify the telephone users into classes in order to identify if a call

does not correspond to the patterns of a specific user; and (3) a

distributed object technique is used for the implementation of this

distributed system (i.e., manager and agents)

Conference Paper Security management against cloning mobile phones

ANN allows you to face the problem without much knowledge about it, so it can be used as a first approach. If the resulted ANN model is simple (small number of neurons and hidden layers) sugest that a simpler and more informative model can be used to solve the problem. If the resulted model is complex, this suggest that you should improve your knowledge domain before adress the analysis.

Very suitable for inverse modeling when the numerical relations between input and output variables are unknown, and can't be established.

ANNs are most suitable for problems that are difficult to specify mathematically. ANNs solve problems by working with examples which makes them very attractive for applications in domains where explicit knowledge is not available, for example, image recognition, speech recognition, natural language precessing, etc.

May this link gives quick understanding,

http://www.cse.msstate.edu/~hansen/classes/AIspring04/slides/expertsystem.pdf

Thanks for all the replies.

# Before using any AI technique to represent a nonlinear relationship between a set of inputs and another set of outputs, you need to discover whether AI is going to do it for you or not. If you as a human can see or feel such relationship by your natural intelligence, be sure AI will do it.

# Be careful on developing your training dataset. A relationship must exist between inputs and outputs, and data points must be enough for ANN to realize such relationship.

# ANN is a multi-input multi-output system with reasonable size of modifiable parameters (as compared to ANFIS, for example). This makes the computational burden lesser during training.

# The main disadvantage of ANN is that you, as a user or developer, don't have much to do to change the behavior of the ANN itself. ANN developing is mainly a trial and error process. All what you can do is to try different structures, activation functions, or training algorithms till it appears to be working.

Good luck.

the ANN have significant ability to derive meaning from complicated data, can be used to extract patterns and identify trends that are too complex to be noticed by either humans or other computer techniques.

The data are feed and verify the result by ANN. The outcome can be transformed to an Excel spread sheet to make the estimation of results quick and easy to apply for any user. This saves both time and money at calculation. A parametric study of the influence of each input measure on the output can done to clarify their effects on the results calculations.

ANN allows you to face the problem without much knowledge about it and ANN is most suitable for problems that are difficult to specify mathematically.

Dear all.

It too has strong mathematical foundation behind it. ANN cannot be separated from mathematics as it more related to the fundamental concepts of random assignment of weights in its functions.

Greetings,

I would like to emphasize the following conditions and advantage:

1) Usage of Artificial Neural Network may have advantage only if you need to detect damage of specific surfaces. It is practically impossible to develop universal ANN.

2) Obviously, ANN may have advantage only if a problem can not be solved by clear algorithm.

3) Under the conditons mentioned above, the main advantage of ANN is ability to be trained. In general case, the wider and deeper a training dataset - the more precise answers from ANN after training.

Major advantage of using ANN over other statistical techniques is its capability to capture non-linear relationship among concerned variables.

I agree with Alessandro; if you do not care much about the meaning of the box, ANNs have powerful mechanism to capture patterns in data. Training them requires understanding the trade off between over fitting the data versus prediction power. In some situations, for example, when you have too many categorical variables, other simple procedures such as decision trees can do a better job.

In a ddition to the above, ANNs are good for noisy datasets.

ANN is nonlinear model that is easy to use and understand compared to statistical methods. ANN is non-paramateric model while most of statistical methods are parametric model that need higher background of statistic. ANN with Back propagation (BP) learning algorithm is widely used in solving various classification and forecasting problems. Even though BP convergence is slow but it is guranteed. However, ANN is black box learning approach, cannot inteprete relationship between input and output and cannot deal with uncertainties. To overcome this several approches have been combined with ANN such as feature selection and etc.

Meanwhile Fuzzy is quite good in handling uncertainties and can inteprete relationship between i/o by producing rules. Therefore, to increase the capability of Fuzzy and ANN, hybridization of ANN and fuzzy is usually implemented.

 One of the most advantage is ANNs allow the modelling of physical phenomena in complex systems without requiring explicit mathematical representations or without requiring exhaustive experiments. ANNs can predict the desired output of a system when enough experimental data is provided.      

Major advantage of using ANN is non-paramateric model while most of statistical methods are parametric model that need higher background of statistic. ANN is black box learning approach, cannot inteprete relationship between input and output and cannot deal with uncertainties. 

Besides the above comments, the use of ANN is quite easy but requires a large amount of data related to the input parameters to cover a wide range of possibilities to predict the output.

 Everything depends on the problem at hand, such as the database dimension, the quality of your data and how much priori knowledge do you have about the problem. If you do not know much, ANN could be a first guess, if the final model is quite small, sure you can use a simpler method to solve the problem. However, much work should be done in order to feed the ANN with clean data. If you have  a laaarge data set, and the problem is a dificult one, then ANN could give you an answer, think about the AlphaGO machine which is based on ANN by a deep learning paradigm.

ANN are adaptive to complex problems, by changing the network topology it would be able to handle different level of complexity 

ANNs are adaptable to solving diverse problems in the field of construction economics and management, good for construction input estimating provided the input and output data are clearly mapped for the software. No previous sophisticated knowledge of mathematics and statistics are required of interested users.