 Good afternoon everybody, I am going to provide some application part of genetic programming which is widely used for now, rather I have used for coastal and ocean problems. Number of problems are studied using GP and then relatively it is compared with ANN approach. This is short outline of presentation, some problems based on special mapping of monthly maximum wave heights, then storm surge problems. Basically for storm surge this we carried out studied based on cause effect modeling and then real time forecasting, special mapping also and if time permits then I will say something about tsunami water levels which is first time used by using this technique of GP and ANN. Now, the general motivation behind this is ANN is widely used for about 15 years for various ocean engineering problems and other problems related to water resources. But there are very few applications reported by GP, specifically few applications are there for in water resources but there are no applications supported in ocean engineering. So there was need to develop some alternate approach which may be alternate to numerical methods also and recently this ANN has been successfully used for so many problems in ocean engineering. Hence, we are presenting here some studies related to ocean engineering and specific use of genetic programming. The problems studied here are the special mapping of storm surge, wind wave heights and tsunamis, then real time forecasting of storm surges, prediction of storm surge through cause effect modeling and some of the studies are listed here which I am not going to present here but work has been done on these studies also. First thing is special mapping of wave heights rather wave heights. Now measurement of wind generated waves is normally done through wave rudder boys or floating boys and this floating boys looks like this where this floating boys transport all the wave information to the satellite and satellites to the base stations that way this information is available to us. Around India there are so many wave rudder boys stations deployed by various agencies and this has been developed all over the world particularly in US also and nowadays we are information is available online through US agencies. Now the first problem here is the special mapping of wave heights. Now this problem pertains to Gulf of Maine, basically when we carry out such type of studies wave rudder boys or these wave heights are very much essential for various operation coastal operations or offshore operations. Now when wave rudder boys get distracted it is very much difficult to collect data at that same locations. Hence we use this technique of GP to prepare a network or to develop a network GP based network where we can hinkast the wave heights of using larger data available as at some locations where the duration is large and the target station where duration is small. For that purpose we choose the station of 4400 this station as a target station where data available is only for around 10 years and for other station the data available was around for 27 years. Hence we develop a GP model to hinkast data at this stations. Hence we call this as a target station for this training we trained GP model on first 20 months of data and the testing was done on remaining 60 months of data. These are the distances available from the coast for these bio locations. Now this study is specifically for northeast of Gulf of Maine where we observed tremendous variations in water levels. Now we prepare similar kind of models using ANN also but I have presented here only the GP results. Now in this case you can observe here this is the time series plot and we could able to model this wave height up till around 7 meters of height and exactly this model has GP has given very good results that R is about 0.98 and the errors are also less. Equivalent comparison studies are carried out using ANN also with network of 4 by 6, 4 inputs, then 6 hidden neurons and 1 as output. So we got these results around 0.96 so compared to this ANN study we got the GP results more better. Similar studies for other locations to validate this technique has been carried out and we found that the results are very much encouraging. Then we move to upon another type of problem storm surge prediction. Now we call it as a storm surge, cyclone or hurricanes. Now basically for this type of study we choose the area of Gulf of Mexico where you observe more variations in water levels. So basically in short this is storm surge is nothing but rise in water level cause of heavy winds pushing water along this. Now this figure is self-explanatory you can observe here how much this water can get pile up to the coast. This is another figure where you can observe much more destruction happens because of this type of events and this is another damage cause due to such type of events. Now basically as I told you these are the three type of studies I am carrying out. The first one is prediction of storm surge based on the cause effect modeling. Now this figure is self-explanatory other 2005 hurricane season. You can observe here number of hurricanes occurring in this region. This is 2004 hurricane region number of storms occurring in this region. So this region is very much own for such type of events. That means people have to model exactly the water level they have to predict exact level and estimate the appropriate water levels. So there are many models deployed in this area but till application of this type of sub computing tool is always welcome. Now specifically this study is carried out at particular locations here. This point 42001, 42003, 039, deep water and shallow water region. Now initially what I did because there are no reported applications on this. As far as NN is concerned there are two or three applications based on the storm surges but for GP there are no applications at all. So initially I took a single event as a catriona because normally these events last for 7 to 10 days. High water level sustained it for 7 to 8 days and then prediction is very much important for that. So initially I took a single event then model training and testing for the single event. Then I increased observing results of this. I increased number of events for training. I took two events for training and one for testing and then again I used hurricane season spanning over 2003 to 2005. So 2003, 2004 used for training. Sometimes 2003, 2005 for training and remaining entire zone for testing. Now as far as this particular area is concerned the highest water level recorded in this area is around 14.1 meter. So what 14.1 meter is huge and we could model our GP or GP could model this height exactly with 10 to 11 meters. Rather when because exactly that height is that water level is not recorded at this particular station. So coming to this problem these are the results because I am not shown here the results obtained for a single event. These are the results obtained using GP and when the GP or model was trained on two hurricane event and tested on one single hurricane event that is named as a catriona at station 42001. Here hourly data was taken into account. Causes to factors like wind speed, wind direction, wind gusts and barometric pressure were taken as input and the rise in water level were taken as an output to the model. So equivalent ANN predictions can show here this correlation as 0.95 and errors compared to this GP as slightly higher. Then another study for cause effect mapping at two stations has been carried out and here we modeled the hurricane seasons for training for 2003 or 2004 and testing on another one. This is a typical testing or this scatter diagram for testing hurricane season on 2004 and model performance is showing good here around correlation 0.90. So this is on improving side rather we are validating our results obtained by this new technology or new technique. This is typical time series plot we can see here this is complete entire data that was given for the training of the model and this was the unseen data on unseen data and we tested model on this particular data. So here we can observe that the water level is around 12 meter and it could model in training also equally good. Also in testing this is typical hurricane A1 where this water level was around 8.77 meter and this could model by GP technique around 8 meter. So we can see that this estimation is going to more and more approximation. This is again a scatter plot for the same. Now recently I am just carrying out one study to incorporate the unseen observed or unobserved parameters at the station because normally we took the data related to the where all observations are taken into account and then when you carry out sensitive analysis then we can arrive for a particular cause effect parameters. But sometimes some parameters are unobserved cannot be recorded at the particular station. So in that case I am developing one model GP based model at one station this is result for that one taking into account wind speed, wind direction, wind gust, parametric pressure as other four parameters usual format of parameters used in earlier study and then observe previous heights or previous wave heights of 1 hour before, 2 hour before, 3 hour before and 4 hour before. I found that while training and testing the results went up to correlation 0.99. So this is somewhat encouraging results by using GP approach. Now then another problem is real time forecasting of storm surges. So estimation is carried out by some numerical method can be carried over some numerical methods also but there is yet to develop some system which can exactly means predict the values or water levels at some particular station during this events. Now these models are developed for the events means when a hurricane occurs and when level rising levels sustained for that particular period how much will be the water level after 1 hour, 2 hour, 3 hour, 4 hour so that it will be easy for the offshore operations to know exactly the water level after 6 hour, 12 hours like this. Here exactly we could model the water levels up till 12 hours only because after 12 hours it was not possible to model the things. For this also we choose stations, deep water, shallow water and near post locations. Input for GP or ANN is given here as two preceding observations, two preceding observation of two preceding water levels. Now this we arrived to the conclusion after so many trial and errors. So this is a typical graph. Here again the two events are tested and training was done on three events. This we can observe here after 6 hours also it is giving exactly near about similar prediction. Sometimes over prediction by using this technique also can be accepted and this is on the positive side. This is another station one hour ahead prediction. Another study has been carried out using this three layered feet forward back propagation training network and we got the similar results. This is for other station 6 hours. Here we can observe that in fact the predicted wave height or storm level is around 9 meters. So GP could predict this water level around 9 meters also. This is compared to study using both the models GP and ANN. Here we observed that up till 6 hours this ANN and GP could predict water levels or the storm levels quite good or GP is having higher age. But after 6 hours we found there is drastic drop down in correlation and there is increase in error measurement compared to GP. Hence we can say that this GP model can predict water levels this storm water levels good up till 12 hours. After 12 hours we tried to model or to forecast the levels but the results rather the forecast was not proper not exactly proper. So we stop up till 12 hours only. Then another exercise or this analysis special mapping of wave heights now special mapping of wave heights means when now this is the storm track for typical RETA event this is for Katrina event. Now when this storm track travels or rather gets close to the coast we have to find out the time from station suppose 42001, 42001, 019 and 4205 and using input of water levels at 42001, 42001, 019 we can or we predicted the water levels at 42035. So this is some sort of special mapping of storm surge but here we should know the exact time to reach the event or that particular hurricane from one station to another station. So these are again results which are very good around this RAS 0.99 and errors are also very less. This is another interesting problem, tsunami water level there is only one application as far as tsunami water level is concerned not exactly to water level but to for what I can say travel type map which is recently published in one of the journal but there are no applications as far as ANN and GP is concerned for such type of problem and this is very complex problem. So we just try to model this phenomena or we just try to predict the tsunami water levels using this special concept. In the sense now we choose the study area where the this is Alaskan area and this is Honolulu point one of the point and this area is prone for the tsunami earthquake. So most of the tsunami occurs over here in this area and propagates in different directions along the coast and this is Honolulu rather it is like. So the first event we took for June 10 1996 event and the earthquake was around 7.9 of magnitude and another location another event was March 26 1964 which occurred in the same region and we studied propagation towards the coast and towards this Honolulu harbor rather. We choose the or the water levels were recorded at this different location these are the bottom pressure recorders water levels were recorded at this particular point and again the water levels were recorded in this region. So after making some sensitivity analysis we use the input water levels at these three stations 71, 72, 73 and predicted this water levels at this particular point where wave propagates because travel time maps are available for all coastline areas in the world. So waves here propagates after 4 and half hours so exactly model by using GP and ANN. Now here we can observe that now the first part the first part of the study was with inclusion without inclusion of tide levels because tides were removed and with inclusion of tide levels what exactly the models can predict because it was rather important to see how the model behaves when the tides are removed and how model behaves when the tides are included. Now this is one of the case of March 1996 where prediction of water levels was made at this Honolulu and the target that was the target station and initially we took water levels at this particular point three points only after that then for another event we took this one we got these results GP predicted R up till point 8 then ANN very less 0.49 and then we coupled the models whatever we got the outcome of ANN model that again model into GP and then we saw how much is the increase in results or correlations or predictions so that went up till point 86 and substantially the errors are also slightly reduced. Also this study linear regression study is made but mostly we compared these models over here this is the typical time series plot and scattered plot here we can observe that it can or it could catch a trend for this extreme event also. Then another was case was with inclusion of tide and for with inclusion of tide we took the another target station of coastal location here because we modeled this study as per availability of the data and it is very much difficult to get water levels at all the stations as far as this event is concerned. So we took this data database from net and again we modeled there are input are slightly changed here Sitka and I quoted as in shown in earlier case we got this result of point 84 here point 71 67 compute here this model could not predict well combined model rather hybrid model but GP could predict well with inclusion of tides this is another event similar conclusions have been drawn now based on these 4 problems or other 3 problems for special mapping of wave heights then tsunami water level predictions for as a problem of special mapping and this Tom surges real time forecasting and then cause effect modeling and special mapping these are the conclusions application of GP to tackle problems in coastal ocean are not as common as the other technique like NNRGA as we tried to fill this wide then estimation of storm surges wave heights and tsunami water levels by establishing special correlation temporal correlations and through cause effect modeling have been successfully carried out and in general GP can handle all these tasks satisfactorily its prediction capability many types exceeded that of ANN sometimes we have seen the results of ANN and GP are similar but in all the cases whatever studies studied I carried out I found that the error measures as for as GP is concerned are always less compared to ANN and in some situations where single GP or ANN model does not work properly then we can combine outcome of one model as a input to another model and then we can get best possible results also this was the first try to apply GP for so much complex problems because as we see the Gulf of Mexico and number of events happening are occurring in that area now it was very much difficult to model that event using this type of technique so here GP could satisfactorily learn the complex wave propagation as we can we have seen already by some examples GP carried out auto regression well and predicted storm surge level based on a given surge time series up to 12 hour sort of lead time and the other listed parameters are already covered here thank you that depth of water rather depth of water at that station location is given sometimes when you change the data or when I didn't get it from when we feed the input data for ANN model it is giving some kind of output and same input is given after a different time period it may not be same result no it doesn't happen see I don't know you have not exactly understood question but what we are doing every time we are giving a sequence of preceding values and asking the ANN or GP to identify that sequence and continue identifying and forecast the next value so as the sequence changes the forecast value will also change every time it will also change every time the subcomputing techniques target the tolerance of real world so what are these techniques ANN and pathologic and genetic these are only the subcomputing techniques or tools or can we include LP and LP into this see that's why you know I said that the exact definitions are very difficult to give basically you know traditionally soft computing techniques are identified with functioning of human brain people have taken inspiration from the way in which our brain thinks or it does cognition and that process is is imitated but these are actually strictly speaking the AWPR and you know support vector machine etc they shouldn't they may not come under soft computing but in general they come under machine learning or data mining techniques or artificial intelligence technique traditionally they may not fall under soft computing they are more they are more to do with developing you know some kind of a artificial learning or artificial intelligence process rather than you know going by the traditional soft computing yes it can you see as an as he was mentioning it can be used for a variety of purposes what you are saying is probably to work out a causal relationship given the state of input causative parameters you have an output or result or effect and you can match those things using then other things are as he's suggesting okay for a given a sequence of previous observations for the next value then you have this what a spatial relationship given the observations of value the different locations to find out the value or some unknown location then as I mentioned the property integration property resolution get all various techniques in which you can find the position of yes there is you know various techniques that was regression enlgp so it was like regression was given quite low regression option was 0.77 for this data so I just wanted to mention like in case of regression there are sometimes there are outliers which it will consider into its modern formulation whereas and might and normally removes those are there is a pre-processing of that up is it there most of the studies did not evolve the pre-processing the way in which you are I mean suggesting yes we don't purge any information you know we don't purge any information we we hadn't pre-processing only in the sense that we have categorized the data into monsoon non-monsoon and then you develop models separately or we have you know filled up the gaps make the data more consistent and use we have not removed any observations as you have seen with outliers but I understand the point you know regression techniques normally great read of outliers first and then fit the relationship in enlgp what happens you know we normally normalize the data into 0 to 1 that it's just maybe taking care of this but there is no you know mandatory kind of pre-processing see this is what he was showing you know there is no straightforward answer to this but the isolated studies done by these two people have shown that gp has a tendency to handle noise in a better way compared to the end isn't it yes in fact for small amount of data also I am going through that study but in that study I found that for small amount of data in fact for training artistic we are making divisions like 70 30 60 40 80 20 like this but when it comes to an n and when when you have a small amount of data an n is not giving the same kind of results for this type of division but if you look at the gp because I have means carrying out number of problem this type of study on number of problems what I studied here so I found that even if you do gp as 80 20 or 60 40 or 70 30 the performance of the model will remain same and an n is best giving results in 70 30 data division in other words he's saying that gp have it has a tendency to work with less amount of data equally satisfactory that's what he's saying if you have if you reduce a small number of training data say gp can work better but again you know these are very specific observations I mean it should not be taken as a gospel truth they may change depending upon the data but our general observations are like this and as you guys I mentioned if the data are too noisy that if there are two random random variations then our experience is that gp works better empty I don't think we have that this data was noisy one for as far as the storm surge predictions were considered because these locations were deep water then near coastal sand near coastal we'll find so much variations in all the parameters this is how it's works like frequent waves yes can it be this outlier right it yes the free waves occur which are very rare but they're very dangerous because of some internal oceanographic phenomena they occur but if free waves occur suddenly there will be a jump in the value and one has to see how it will never walk on the number it assays but if it happens then we have to say see whether why Google of those outlines the results still in our case should we ask about the remote centers anyway I am always there