 So, in continuing with our discussion on NN, I can just once again want to stress this point that you have to apply the neural networks where they are really needed and the problem should be such that the application of neural network should be justified and there are no satisfactory solutions by the traditional techniques. For example, one work done by one researcher is like this that ISRO collects a lot of data on satellites like Topex, C-Sat, etc. Now, when they collect the satellite data in such a way that the satellites evolves and comes to the same station after every 10 days or 11 days and like that and like that there are so many repetitious paths that are there. Now, the values that are sensed by satellites are reported in such a way that they are suppose this is a path of a satellite or a track of satellite, then at every certain distance like 1 degree by 1 degree or so, the values that are sensed are averaged out in both time as well as space domain and then they are reported. Now, a problem comes when you are interested in only some station specific value especially to coastal station because the reporting at locations which are very near the coast is something always suffering from some flaw because there are very few observations when the satellites go near the coastline. So, the coast near coast data are not reliable therefore, we wanted to devise a method by which we can translate this deep water data of wave height, wave period, wind speed to some shallow water location. Now, traditionally this is done by this numerical methods, but numerical methods have their own problem not only apart from complexity of the numerical scheme but they have modeling difficulties there are also certain numerical constraints especially when we go to the near shore data. Therefore, we thought to provide a technical to an alternative to the existing one to neural networks and to project the deep water data to shallow water. Now, so we had this data collected by satellite topics of wind speed, wave height, wave period. We also had for calibration purpose the data collected by instrument called wave rudder boy at these locations for 4 years. Then we developed a neural network wherein the input was the data of wave height, wave period, wind speed at these locations several locations and the output was the resulting wave height at a shallow water location. So, this way first of all we saw whether the projection of one parameter alone is sufficient. For example, we will take the data at let us say about 21 stations separated by one degree and projected to the required coastal location is W3 near Ratnagiri. And second time we developed a full network wherein we projected all the data collected at these locations and obtained the outcome of wave height, wave period and wind speed at the shallow water location. Then we find that you know this network which brings in lot of flexibility by virtue of largest size what was necessary and this proves to be better. Now, we also separated the data as per monsoon and non-monsoon considering the variability in the statistical parameters in these two space of components and we found that if we do this preprocessing the results improve. So, this shows how far the output of wave height, wave period and wind speed during monsoon, non-monsoon season resembles the actual observed values in terms of correlation coefficient mean absolute error and root mean square error. And this shows again the comparison between the time series comparison corresponding scatter plot in which we have compared the observed significant wave height with the network derived, the observed wave period with the network derived period and the observed wind speed with the network derived wind speed. And we found that the choice of appropriate network gives you the correct values. Now, the point that I want to stress out of this presenting the studies like this this that this study showed how the wind speed and periods along with wave height sense by satellite in deep region can be used to derive the transport values over a specified coastal region using ANN. Now, we found that if we incorporate statistical homogeneity in the measured values in training then it can efficiently tackle the highly random variations of input in that way. Although preprocessing is not a precondition, then we also found that the larger the difference between the such statistical variations in between two data sets monsoon and non-monsoon the better is the gain in the accuracy compared with the unsaggregated learning. And most of the earlier study is they could not translate the wind speed properly they used to get very high errors when they wanted to translate the wind speed. But in this study by proper choice of input output scheme and by proper choice of parameters control network control parameters we were able to get very good accurate results for wind speed also. Now, the importance of the studies like this that we require instrumental data at both at the target location only for calibrating the network. After that calibration is done you can take out that instrument deploy it elsewhere because collection of the instrumented instrumental data is indeed very very costly. So by developing a technique we allowed the data collection agency namely ISRO to collect data for these locations of interest only for a small time and use that instrument subsequent place so that they can save lot of money time and effort in collecting data at many locations. So like that there are several advantages of using the technique of neural network. Now so in the end therefore in the end therefore that I wanted to stress that to summarize what we have seen just now we have to know that we have to apply the technique of neural networks only in the cases where its application is justified. The problem should be really such that it is indeed very complex and secondly the available traditional methods do not give a very satisfactory kind of answer. In that case only you have to see application of neural networks even in that case first you try to apply the existing methods and see what level of accuracy you get then you apply and then see whether there is a substantial increase. Now sometimes people I found people have a tendency especially the young researchers that using the traditional say numerical methods or statistical technique I get let us say correlation coefficient of in between the observation and the estimated value something like 0.8 and by using A and N I get 0.83. So that difference is quite marginal it is not that for the small difference you use this neural network nobody will accept the neural network your neural network formulation because you give them very small improvement with the existing. But if the existing statistical performance criteria differ much much than the neural network based criteria then people will accept your results. Secondly you just do not stop by giving the network you ensure that it works in practice carry out the parametric studies and see whether it really resembles the physical process or not if not you have to try some alternative network architectures training schemes and training forms and also when you do parametric study there is also a scope to understand the physics in a alternative way or in a better way. Now discover the new parametric relationship possible using A and N that can also be done another thing is that always try to decipher the black box nature of the network as we have seen and finally we have to see that at like any at in any other disciplines in our discipline of water also people are now advocating use of hybrid models in place of the ordinary single type of modeling.