I have used this algorithm, effects was quite good. Some info you can find in my papers about making music with ANN. But it is necessary to remeber that using methods depends on problem being under solving.

Resilient propagation(RPROP) method is an optimizer for artificial neural networks. The acceptability of a method in machine learning is dependent on error surface of the training data. If error surface contains lot of plateaus, then RPROP is the best optimizer as per my knowledge. But, if error surface contain lot of ups and downs, then RPROP method hardly helps in improving neural network.

On some data-sets, it even outperformed momentum method. As it is difficult to imagine the plot of error surface, the only practical way to judge whether to use it or not is to train the network and check test error. For theoretical details of RPROP method, you can watch geoffrey hinton lectures on coursera.org.

I suggest you to try the two papers at

http://www.ijer.ir/?_action=article&keywords=ann

Dear Satya narayana Vusirikala as you said "It is difficult to imagine the plot of error surface" I also think similarly, but my network was trained and generallized well. According to yours and Wojciech Zabierowski  reply it seems my results are authentic!

Thanks A. R. Karbassi for suggesting these articles, which are helpful too.

Resilient BP is as good as other methods, only changes the way the weights are adjusted, that can be better in some problems than others. For example, Levenberg-Marquardt is very good for error surfaces full of mountains and valleys, but it tends to get stuck if there exists contradictory data (equal input patterns are associated with different output labels).

The worst approach in your experiment is that training and evaluation with so few data is dangerous (84 data pairs). You should try to implement a N-fold cross validation strategy, where N networks are trained with different fractions of data. For example with N=7, you divide the data in 7 sets (each one with 12 random selected pairs), and each time use one of these sets to evaluate and the rest to train the corresponding network. The final resulting model is obtained with the mean response of the N networks, so the confidence in the result is increased.

Thanks Julio David Buldain Pérez for your suggestions, I ahve already successfully trained ANN by Trainrp as well as trainbr (I trust more on trainbr due to automated Bayesian regularization). Suggestion for cross validation is also good, I hope this will work too. But I don't think 84 data pairs are not enough, as it is an engineering problem with 3 inputs with 7,6 & 2 factor levels, further increase in factor level is not possible.

Yes, if the problem has a low-dimensional input space (7x6x2=84 input states), then it is not necessary many data to represent the space. However if noise is present in the input values (for being real numbers instead of discrete values) it would be necessary to have more measures of each input state.

Perhaps your problem is more like a look-up-table, I don´t have clear enough the kind of your problem: ¿classification or approximation (regression)?

Thanks Prof Julio for your suggestions, it  is an approximation problem. todate the results are satisfactory. MAE error is max 2 on scale of 30-100 with R2 between 0.95-0.99. But on 10 % cross validation few sets have MAE 4 on scale 30-100 (output values) and its mean is around 2.3! Can I randomize the data to reduce the error on those few sets?

In my experience, for approximation problems, the best training model is Bayesian Regularization, as this method tends to produce null weights in all irrelevant inputs, so the approximation model becomes less affected when those input variables change. Levenberg-Marquardt is not so good (always supposing that the output surface approximated is not too much undulate) and needs that each cross-validation set must train a pool of networks (say P networks), so the final model has NxP networks. Bayesian regularization usually replicates the same weights for a certain data set, so they only need N networks in the model (one for each cross validation set).

If the error is high in one validation set, it is caused by the particular selection of samples in the training-evaluation sets: a few bad or difficult data samples had appeared in the evaluation set, and the training set has not enough information to cope with their output. You can repeat the model changing the number of cross validations, or try randomizing the data and repeating the cross-validation set separation. By repeating the process generating several models to average their response, the model becomes more robust. As the problem is quite small, there is no problem in this strategy, where the final model is the average of many models (say T of them) from N-fold cross validations, so it contains TxN networks.

I was not lucky with the Matlab results at all. In my case, another software developed only for ANN and results were fine.

Yahya A Asiri thanks for sharing your experience. But I used to train ANN in Matlab using own modified GUI.

Please, contact Dor Catalysis fr. Martin Holena, Leibniz Institute for Catalysis Rostock, Germany. He is an informatic scientist from Pague.

Manfred

According my experiences, the RPROP algorithm plays an important rule in optimizing ANN models. Using criteria validations like RMSE, MAE and MAE show that RPROP not only accelerates the learning process but also it helps to achieve fascinating results for me.

Interesting...