 Yes, I will just show you one study wherein we have applied ANN and GP and some very interesting conclusions probably can also be seen out of that. The study pertains to application of ANN and GP to estimate the coastal sediment transport. Now as some of you may be aware that the waves generated by the wind which is a continuously occurring phenomenon in the sea, they break when they come to shore because they cannot sustain their motion in shallow waters. Because of that a current is produced which has a component parallel to the coastline. Now this current carries lot of sediment in it. Now when such sediment movement which is parallel to the coast crosses a deep channel which is used for the ships to navigate inside and outside the harbour, they deposit that sediment load, lot of material gets deposited and ports and harbours they spend tremendous amount of time every year in dredging out that material. In fact in India we spend I do not know maybe more than couple of thousands of crores of rupees per year in all these ports and harbours to dredge out that material. So the estimation of that sediment that gets deposited is indeed necessary so that in port budgets you can keep that much money for dredging in the next year and so on. So accurate estimation of siltation is indeed necessary and but this is very difficult because the phenomena involved are indeed very complex. Now currently what is done there are only empirical formulae which are popularly used to estimate that sediment transport and therefore we try to see whether we can get good results by GP and ANN. But as I mentioned ANN cannot be just applied like that because it suffers from some problems like lack of guarantee of success in a new problem, new application. Accuracy levels obtained are arbitrary and you have to go on perfecting it. There are also a difficult choice of training schemes, architecture, the learning algorithms, very control parameters etc. And therefore any new application of ANN must address this issue and then present the results to the scientific community. Now therefore in this project our objective was to develop a suitable neural network to obtain the longshore sediment transport rate LSTR, Longshore Sediment Transport Rate. We also wanted to understand the processing of the input done by ANN and done by regression and also physical consistency. We also wanted to develop estimation procedures on other soft tools like GP and if necessary we wanted to explore possibility of developing hybrid models. Now as I mentioned traditionally there are certain formulae which the design codes recommend. One is such regression formula given by coastal engineering research manual, another one is Walter and Bruno equation, third one is energy plus method. They all relate the outcome to the causative variables in an empirical manner. Now we have collected data, one of the co-authors of this paper collected data along the coast of Karwar here, this is that Arge beach, there is a 4 km long coast and in that coast in the field for about 4 months they collected, he collected the data from February 90 to May 90 of significant wave heights, average wave period, breaking wave height, breaking wave angle, width of surf zone, mean sediment diameter, long shore current and so on. And of course the sediment was also collected in order to know its rate Q per unit time. This is a typical mesh trap to collect this sediment. Now we had to discard a large amount of data because all these values were not simultaneously available at many times and ultimately we ended up with only 81 data examples. Now note that these 81 examples are looked short but in field it is very, very difficult to collect all these variables by means of instruments which are indeed very costly and very difficult to operate simultaneously. Now to begin with we developed a simple network wherein the input was the causative parameters, the output was the rate of sediment transport. It gave us fairly good measure, good kind of performance during testing that means with respect to those data sets which were not seen by the network in training. This is the predicted rate of sediment transport, this is the observed rate of sediment transport. In general the correlation coefficient and mean absolute error and root mean square error were quite fair. Then we all based on the data set that we newly obtained, we also fitted some regression equations like this. And when we compared all these results we found that by manipulating with the networks properly we can get a good network having fairly good amount of performance criteria. Correlation coefficient root mean square error and mean absolute error compared to the multiple linear regression and non-linear regression of different types. But of course within the neural network itself certain training schemes can give you much better results. Now again it is not, there is no guarantee that use of some particular network is going to give, always give you a better kind of result. You have to try out all these training schemes in network architecture. But the point that is to be noted is something like this that when we compared the values of rate of sediment transport predicted by formally advocated by design codes and the actual rate of sediment transport we got this picture where we expected all values to fall on this along this line of exact fit. But if we use one formula RCE, RC formula this is the picture that emerges, if we use another formula this is the picture that emerges in fact certain negative values are also predicted meaning that the drift is in the different direction, negative direction. So like that first of all we ensured that the existing formulae at the given location of that car var are definitely not good. You have to go by the new formulations, new formulations could be neural network because neural network is substantially high level of accuracy compared to these existing ones. Then so accordingly we recommended the network then we also ensured as I mentioned that the neural networks behaves in a consistent manner with the physics of the phenomenon as the wave height increases the rates of this sediment transport also the increases. Now yes now then another study that we did is like this that we did a parametric variation with respect to traditional regression and also parametric variation with respect to neural network and that gave us some insight as to why regressions fail and why neural networks excel. As you can see here in this particular figure wherein we have plotted these sediment transport rate with significant wave height both process by ANM and by regression and in this next figure we are we have plotted the rate of sediment transport against the increasing breaking wave height and using the ANM and regression and in both cases we are finding that when it comes to regression you know the regression tries to do some kind of very rigid approximation to the input whereas here there is a lot of flexibility in modeling. So neural networks tries to I mean the traditional regression tries to pass the data through some very constricted path whereas lot of degrees of freedom are available when it comes to the neural network. So neural networks invariably enjoy what we call a high degrees of freedom in processing and that is why they perform better. So similar conclusions were drawn by other variables yes then another thing that we did is like this neural networks gave better results than regression but even then it is still possible to go at a higher level. Then what we did we developed a basic network using the training set of data then we repeated the training by attaching another network in series found to it. So the idea was that this network will do the basic function approximation and fine tuning can be done by another network later because by doing by giving a network like this you bring an additional you know flexibility to the data modeling and that work again very nice the correlation coefficient again substantially improved compared to the single stage neural network. Then many times you know these are these things are used by field engineers if you talk about neural networks etc they may not be very receptive. So what we did considering the fact that some this non-linear regression results may also be acceptable to some level we try to give a coupled equation that combines the input of the regression and then computes the output with the help of neural network and such an equation based on a coupled form is given over here and that also works fine but not as fine as the pure neural network or two stage neural network. Then we applied the genetic programming and we found that it gives results similar to a good well trained neural network this is the graph of predicted sediment transport rate with the observed rate predicted was calculated using the genetic programming. Then again we did so many things to see if we end up with some better results we trained the model using ANN and results of that trained model were used as input to develop a genetic programming and obtain the output of the rate of sediment transport accordingly and these are the results first we did training using GP primary training and fine-tuning using ANN here we did find primary training using ANN and fine-tuning by GP and we found that all these methods also give results which are definitely I mean they are not an order of magnitude big than greater than the ANN but consistently they were found to be better than a single ANN. So this is what we found so if somebody wants to take a cue from it and applies to and wants to apply the same combined or hybrid modeling in his own case he will find some good elements of his I mean some good hints out of this study. Now this is a locality learning actually they belong to a class of models which are called local models. Now so far this ANN and others we have seen that they are global models in that you know you take the you develop a model with the help of trained data and continuously use the same model here these are local models in the sense that whenever any new input is given they will find out which are the they will pick up the related data from the input database and using some criterion and apply a local model to it and then use it to obtain the output from that new input. So these local learning belongs to that class of models called local models which are different than global models in that there is no single model that fits to the all data. Then the local models fit data into a region which is called around the location of the new input or query point as it is called these are there are certain terminologies which are traditionally used by those mathematicians or you know control theory people. Now this locality learning is a memory based learning in that it does not discard the training data like ANN but you make use of every data or part of data every time in making a new learning. So learning is rather incremental here it is non-parametric that means from the data you do not work out the parameters once for all and use these parameters but every time you do some kind of incremental learning over and above what you have done so far and locally linear models are selected and fitted. Now so basically what it is done is like this see suppose this is the new input x then in locally weighted regression we using some kernel function like the Gaussian function as you can see here is a Gaussian kernel function this x is input variable and this e k and d k are the two parameter centers and distance metric distance metric could be typically a Euclidean distance metric roots square root of x equal plus y square. So like that using this kernel functions you this is a typical kernel function shape this is a plan you can say so for a new query point or new input you identify the training patterns which are closed then to those identified local points you fit a regression and use it to obtain the output this is I mean conceptually it is like this then you have a class of locally weighted regression models called locally weighted projection models in which the input data are mathematically projected into these orthogonal or perpendicular directions and along each direction a linear model is fitted after identifying the required points with the help of these kernels and average of all this is ultimately used to obtain the output. So the things that here are indeed extremely complicated and the mathematics involved is quite involved is quite complex to explain but conceptually the things involved are like this. Now so far two of our students have used this technique again to solve similar problems estimation coastal sediment transport rate from the causative parameters and estimation of score depth of a schism okay. So this is the same problem which I was mentioning the data near Karwar was well taken the input in this case was significant wave height period width of surf zone breaking wave height and the output parameter was rate of sediment transport and as you can see this is the predicted versus the observed sediment transport rate this is the exact fit line that means had the data been perfectly predicted all points would lie along this line we got a good correlation coefficient which is 0.88 comparable to usual Nn but again you know we had to go on refine this technique and see whether we it can give results which are more acceptable than Nn or not then of course they work much better than this traditional techniques recommended by design code that is the definitely the advantage. So we found that the locally weighted regression technique they produce more accurate estimation of the rates of sediment coastal sediment transport than the traditional formulae. Another application of this LWPR technique is the estimation of score downstream of a spillway again the same kind of a data observed in the field and models was used the this is the maximum score depth it is to be related to this head then this discharge radius of bucket lip angle etc and as regards field data were concerned there were only measurements of discharge and head used to obtain the depth of score in fact the that US our Indian standard code recommend only use of this Q and H1 to find DS for generalization purpose when we use these two parameters we found that LWPR resourcing about 0.9 correlation coefficient during testing and it works better than the traditional formulae if we use the model data where apart from discharge and head radius mean sediment diameter then lip angle and then the other tell water depth DW where use that input and that significantly you know improved the relationship we got a correlation coefficient of 0.97. So it means that the practice that is followed in design code of using only discharge and head to obtain the score depth is no good you must include other parameters also in order to get a better estimation of the score depth. So this shows the relative comparison between the results of a locally weighted projection regression then the best trend ANN and the formulae used by different investigators in field like Martin Sveroni's who in see a new statistical regression also fitted that also didn't that didn't work much better. So right now we can say that LWPR results are comparable to that of artificial neural networks maybe if we are able to refine this technique further we can they can go one step beyond ANN and also but right now we have to keep the fingers crossed. So we are only saying that this core prediction could substantially improve if the LWPR regulation technique is used in place of the traditional formulae. So like that so many studies have been made with respect to this GP MT and so on.