 Yeah, so I have the pleasure of being able to speak to you as the first today and it's a great pleasure, of course. And according to the program, it seems like I have around two hours for this small presentation, so that was a joke. I don't have two hours. No, I will probably try to keep it a little bit short, but nevertheless, I would like to give you a few, let's say, perspectives on the theoretical framework. And we started working, preparing also a draft paper which will be available, but it is very much a draft and it's something which I suppose that we will develop much further here in the coming months, not least based on the feedback from all the good experts sitting here in this room. So this is a very first contribution by Dimitri. Well, where are you Dimitri? Yeah, it's sitting there. So this is Dimitri Well and Sebastian Tönsch you already have seen and I am Michael Faber. And now I'm trying to see if this would actually work. Yeah, it seems so. Something is working. Okay, so I would like to go through these few points in the next couple of minutes, looking at the underlying motivation. A few words on structural health monitoring, engineering decision making, decision and VOI analysis. And then I will just introduce in a very, very simple manner the basic concepts of VOI analysis according to the Bayesian preposterior decision analysis in the extensive form as an example, whereby we can adjust a little bit our mindsets. Then I'm going to point at some interesting recent works in the field and I will conclude by suggesting a little bit as summing up of the situation as it is and in what directions we could go here in this project to start out with. Okay, well, in terms of managing the integrity of structures, I think it would be fair to say that Freud and Teil was really the front runner and we are talking about, let's say his work in the 40s and we often reference one of his works from 1947. So the fundamental idea that we should establish a rationale based on the available information to base decisions on the management of integrity of structures is definitely not new. And a lot of developments have really taken place since Freud and Teil and it goes beyond my time limitation to talk about all these contributions, but you surely know and appreciate the enormous amount of work which has been performed in the period between then and now. It's then interesting also to see that the decision theory, the Bayesian decision theory in particular has taken some big steps forward, and what was basically developed by Riefer, Riefer and Sliefer sometimes in the 60s and also let's say the fundamentals of utility theory came out from Neumann and Morgenstern and a rationale was let's say crystallizing exactly in what way it would be correct to treat uncertain knowledge in the context of decision making. Now, we should also highlight that Benjamin and Connell actually at a very early stage realized that he was something really important and really big and in their let's say the standard textbook on statistics and probability for engineers, civil engineers, they highlight that and they describe let's say the basic constituents of the Bayesian decision analysis in a very straightforward way in the prior posterior and pre-posterior decision analysis and they even give a lot of good examples, but recognizing how long ago that actually is, it's pretty amazing how little has happened since then. So the application of let's say especially the pre-posterior decision analysis in engineering, in civil engineering maybe in particular, but I'm not sure has been long on the way. So the merits and potential benefits of utilizing pre-posterior decision analysis and value of information analysis has not really been realized and exploited and we are in many ways we are standing right in front of a lot of interesting work in order to exploit this potential and this action is of course meant also to push this in the particular context of structural health monitoring. So I would like to point at let's say a suggestion for different classes of engineering decision problems, civil engineering decision problems where we could utilize pre-posterior decision analysis and value of information analysis. I would also like to give you the idea that management of structural safety in principle can seen as purely information theoretical problem. So in an abstract sense what we are actually dealing with is just managing information and I hope to give you that idea. And then as I already indicated I will show in a very simple way how the value of information analysis can be used in the context of structural health monitoring. So one way of classifying decision situations is what you see there. Structural health monitoring could be utilized in the context of prototype developments. It could also be used in the context of code making and calibration both for design and for assessment of structures. It could be utilized in the broader context of optimizing strategies for warning, warning let's say in the face of hazards or perils which could either be due to let's say emerging natural hazards or due to emerging deterioration or other things where we have a time dependency and then of course for the optimization of maintenance strategies where we normally are focusing. If we look at prototype developments health monitoring of new structural concepts like wind turbines is a very good example I think where we are developing we are trying to push the designs of new concepts all the time. But at some point in time we are fixing things and then these things they go into mass production and they are put up everywhere in the world. So we are talking about many many many structures which are put up in let's say in their environment and in that process prototyping is certainly a useful concept and it's also applied. But the structural health monitoring in order to understand the performance of the structures and the types of loads and their responses. This is also being utilized as it is now but here what we are doing could also contribute in a more rational and strategic manner to optimize design decisions. When we look at code making and code calibration for design and assessment of structures again an important issue concerns the and this is I guess it has over the years become more and more clear to me. That model uncertainties with respect to the performance of structures in reality as compared to the simplistic models and very often very highly idealized models we are using play a very big role. And structural health monitoring can help us in a strategic way at the level of strategic decision making on how to adapt and calibrate codes rationally. Then the issue of optimizing warning measures that as I mentioned is a context where we have had some experience it's really the type of problem which can be cooked down to should I stay or should I go should we do something or should we leave it. And that type of sequential decision making problem comes up in a number of cases in engineering. So if you have a typhoon emerging somewhere and you have some facility or structure you're worried about could be hit but you're not sure that the typhoon will come your way then this is a type of such situation. And let's say the monitoring is on the location and on the characteristics of the of the peril in this case the typhoon. But there are many other similar contexts should we stay or should we run. And then of course for the optimization of maintenance strategies the knowledge which we can collect during the lifetime of of a structure certainly is very useful for supporting decisions. This has really been realized I would say already in the beginning of the 80s there were a couple of groups I believe which can be attributed to significant work. The second work one was a group around pallets of Christensen and John John was a small boy at that time about this height and I was a little bit smaller. But Pelle forced us to work in this domain and at the same time at DNV they were also working in this area. Henrik Madsen and his colleagues and that was then the developments on risk based inspection and maintenance planning which which was born. And that was really a preposterior decision analysis in order to allocate resources optimally for management of integrity of structures. Now structural health monitoring in many ways can add to that because these optimal risk based inspection plans can of course be adapted during the life cycle of the structures based on even more information which can be collected in between inspections. So the fundamental logic is that we sample information knowledge over the lifetime information costs resources money ultimately depending on what technique we utilize. We get more or less precise information but typically there's also the relationship between the precision and the cost. Using this information we can adapt our strategies on how actually to intervene with the structure and this is where we get the potential benefit out of the whole service. Yeah if of course the information we collect is not correct or biased the derived actions which we will identify as being optimal may actually not be optimal. And that again will not provide us optimal solutions. Yeah so the only basis we actually have for evaluating the benefit of collecting information during the life cycle of a structure is information we don't get have. We it's still out there we can we can pick it up when it's there and the benefit of this does is the issue how can we evaluate the benefit of this right. And what has happened typically so far when when the effect of structural health monitoring has been assessed and also quantified mostly is that let's say time series of certain phenomena are anticipated. And then you see OK so if we use structural health monitoring how could this be utilized and which could in that particular case conditional on this particular evolution. What would the benefit of the structural health monitoring then be. But this is of course not correct. We don't know exactly what the future will bring and therefore therefore we need a different rationale. There's no doubt that the structural health monitoring can significantly help in supporting decisions. Very little has been done so far on this and there's no doubt that what we will do in this course action will certainly contribute. When we talk about decision making in in the context of assets management or structural integrity management it's useful to look at at the system is a let's say a system comprised of different constituents. And we are trying to make decisions on the basis of the knowledge we have and we can in principle we can make two types of decisions either we can change the physics of the system or constituents of the system. So we can go in and make actual physical maintenance activities we can repair we can do things and in that way we actually shaping the physics. We do that or we make our choices based on knowledge we have concerning the constituents. So it also makes sense to direct attention to the knowledge we can collect the additional knowledge we can collect about the constituents. And the utilization of this additional knowledge to support a let's say a more optimal ranking of possible decision alternatives concerning physical changes. So it's in principle it's always a mix of these two principally different types of actions we can take which is relevant. And in some cases of course knowledge is already sufficient and additional information will not lead to anything except additional expenditures. In many cases let's say carefully devised and strategically planned collection of additional information will actually provide a benefit and this is where the value of information analysis comes in. If we look at the different let's say places where we can collect information when we are looking at the system and its constituents then of course you can make many models. I've just taken the one from the joint committee on a system representation and as I said other models apply but here at least you get the idea that yes we can monitor the performance of structural systems by looking at different things. So we can look at the exposure side. We can also look at the damages which would be down here on the vulnerability direct consequences but we can also collect useful information on the functionality side the indirect consequences. So information can be collected at different levels and all this type of different information can provide very important knowledge for managing the integrity. And all this could be related or understood as structural health monitoring. Yeah and of course we try to model the scenarios of events for such systems and also to calculate the corresponding risks. Now what I said earlier concerning this perspective on management of information that that actually in many ways at an abstract level really describes what we are trying to do. When we are when we are making decisions with respect to design or with respect to particular maintenance activities design of course involves choices of materials. In addition to the geometry and the statistical systems for instance and also any any other issue of dimensions for structural systems. And when we make those choices what we really do in essence is that we are we are buying we are buying information so when we choose one material for another material we are buying information concerning the performance. And this information we immediately using let's say probability theory and methods of structural reliability we can immediately transform that to a probabilistic description. So already in this phase what we actually buying by designing we are buying knowledge we are buying information and that information goes into the same pot. As any other information which can be collected about the structural systems at a later point in time using for instance structural health monitoring. So all together I think also in this action we should keep an open eye to possible let's say formulations and tools out of information theory. Yeah. In principle any design and also reassessment decision may be supported by prior decision analysis. Any decision on the assessment inspections for monitoring may be supported by previous theory of decision analysis. Now when we are talking about monitoring there is another issue I would like to underline that it's actually it concerns all sorts of collections of information. Normally we understand by monitoring some continuous in time observation of some phenomenon but whether it's continuous or discrete or only one time or two times it doesn't really matter. It's just a broader term for collecting information in the prior decision analysis it's all here. Very simple so we have a space of possible decisions depending on the choice. We will have an outcome of nature and depending on the outcome of nature and our decision there will be an associated benefit or utility. We call it just benefit and well we need to optimize decisions so that we get the the highest expected value of the benefit expected value that's dictated by von Neumann and Morgenstern. The axioms of utility theory and and that's that's very simple. So we buy information by choice of the prior density. This is essentially what we're doing. We select about possible different choices of of prior probabilistic descriptions of this universe. So by buying a cross section by buying a geometry we get information about the the the performance right and then we take this in knowledge into account when we evaluate the the benefit function. Okay so base here and all this is very simple. We if we have additional information which we represent taking into account the uncertainty this information is associated with then the knowledge we can build up the posterior on the basis of these. Information's weighed with our prior on on on on this particular phenomena of interest and this is what we call updating right. And then we can do posterior decision analysis in the same manner as we do the the the prior decision analysis but now we remember here we have the double prime indicating that we are taking the expected value over the. The posterior description probabilistic description otherwise everything is the same. And as I indicated earlier this is the extensive form decision analysis. Which we then come into here where we look at the possibilities to include new information. In decision support even before we have collected this new information and we do that by assuming that this new information that we are seeking will follow our prior models. For the phenomena and then we take into account this information we average out or all possible outcomes according to our prior models. But every time in a way that we have realizations in this integration of the outcomes then we optimize the decisions in the decision problem. And then we do the same again and we average and maximize. The choices on how to collect information so this is what we're doing we are taking decisions on how to collect information we are observing what comes out of this these experiments and on this basis we make decisions. Based on the decisions and everything which has happened before we have a realization of the universe and associated benefits. And this is the extensive form formulation of the previous decision analysis and I just also want to highlight the normal form. They are of course equivalent the major difference in the extensive form normal form concerns the necessity to include decision rules in the normal form decision analysis. And then of course the the big difference operationally is that in the extensive form we operate on the posterior probabilistic description in the normal form we operate solemnly in the prior probabilistic description. And that of course gives some let's say practical differences. And when we are looking at all the early works on risk based inspection planning using the previous theory of decision analysis you will you will also realize that everything is in the normal form. And in the later years developments on value of information analysis and structural health monitoring immediately all the emphasis has been put on the extensive form formulations. I'm just saying that we should keep in mind that there are two equivalent formulations and we can we can utilize those of course to the to support us in finding efficient schemes for taking for doing the analysis. Yeah okay I want to point at a few recent works in the field and I actually have taken all these papers with me. Including one I cannot remember the name of but that was this one. These are some of the more recent works but in addition there's also a very nice let's say a global overview of value of information analysis utilized across let's say all scientific fields. And this is a paper by Jeffrey Keisler and Collier and shoe and Sinatra and link off and it has the title of value of information analysis the state of application. And that that paper of course is is really interesting because it gives a broader overview of the degree to which the different sciences and engineering fields are taking benefit of value of information analysis. These these references which I list here they are selected of course there's more out there and in particular I would like to draw your attention to a couple of PhD thesis which have emerged here over the last four or five years I guess. And one is Timos is sitting down there Timo hand up. Another one is of course by posse a colleague collaborator of Q Regan and another one which does not have a title anywhere close to structural health monitoring but which is really related to solving this sequential decision problem. Should we stay or should we run which is highly related to this is by Annette Anders and then I believe Daniel is your PhD student working in this area done now or. Yeah and I'm sure that that there will be other other ceases emerging or also already available. These these contributions here common for those contributions is that they almost all use the extensive form decision analysis formulation. They look at the in principle different variants of the same problem you have a system. You have time involved. You have the possibility to collect information concerning the the condition of your stock deals component or system because some of them they also go into different types of simple systems. And then you were based on what you collect and how you use this information in order to update your knowledge on certain cities. You can optimize your intervention. So this is really the issue of posse and Q Regan. They look at the identification of let's say efficient efficient strategies for collecting information by looking really at the precision of the of the information you get from different techniques and you can represent this very very nicely in a given context. Santa defines a member of society with the name Tom and he is rational. This is a claim and and Tom is collecting information about the deterioration of the bridge in order to optimize his decision making utilizing a couple of different types of monitoring devices to control. Well the bridge is is deteriorating beyond what is reasonable. Here in Conakley and myself here. Katarina is looking at more or less the same type of problem as posse and Q Regan. But also looking at simple parallel systems and serious systems where we have the possibility to not only collect information about one component but collect information about several components. And then of course depending on the the dependencies between the components and their performances this information can be affecting our probabilistic models of the other components as well. So so Daniel who oops sorry who's with this paper here looks into looks into the the formulation of of the of the equations which we're utilizing in order to calculate the expected values and their pin points that of course instead of just integrating our our weight utility is is is is is as you would be inclined when you see the equations you immediately also realize that this integration can be transformed into a series of of let's say typical and well-known operations using structural reliability theory. So by identifying the relevant limit state functions which are involved and go into the calculation of the expected value of consequences. There seems to well it's it's quite obvious that some significant benefits also in terms of computational efforts can be achieved. And then also some directions on how to achieve efficient numerical treatment of of of these probability integrals using simulation techniques. In Rollscore Joan who's sitting over there the main topic is to utilize forecasting models which are of course associated with a significant uncertainty as a basis for taking action on reducing expected value of losses in the context of falling ice from cable stays on large cable stay bridges. Yeah and in the last paper here which it's not actually published yet so but it will be at the icosar. It's it's really a kind of of of similar work in some ways as the as the approach suggested by by Daniel where the where the calculation of of the expected values the expected benefits are undertaken in terms of structural reliability methods. But then where also a generic model is introduced on how to represent deterioration of of systems of structural components. Sorry not systems is formulated and it's shown in in let's say in in a very simple way how the different operations are to be undertaken in the decision analysis. Right concluding in principle value of information analysis is it's just another way of of of of describing let's say some of the outputs from preposterior decision analysis. So it's that there's there's nothing fundamentally let's say new in value of information analysis. It just comes out of of preposterior decision analysis and it was utilized already heavily and it has been over the last 25 years. And there's a lot of let's say ideas and and approaches which could or could not be very useful also for this project here. The preposterior decision analysis has been applied to a number of different structural health monitoring problems already with success. I think what can be said is that in principle it's straightforward to formulate the decision problems in principle. When it's only in principle it's because that those formulations are hardly to be found anywhere. So it's a sad reality that we are instrumenting structures every day but we actually we we we hardly care to formulate why we do it. Not are we not able to quantify the benefit of it but we are not even able so we don't even care to formulate why are we actually doing it and in what context are we really going to apply the information. In this paper this overall study this paper is relatively new from 2013. Likewise in the conclusions of Daniel's paper it's highlighted that maybe at some point in time someone should try to not only quantify the value of information but also quantify the analysis of value of information in terms of benefit. So what is the actual benefit of going through the efforts of undertaking a value of information analysis? We need simply to provide methods so clear and tools so easy and to disseminate whatever we find here in such a way that these approaches this way of thinking will be much better anchored and distributed. Major challenges of course the computational efforts these event decision trees they generally explode. They become really big very easily and the one shown by Sebastian well it was called generic but I you know in a way yes it is generic but all the different types of decision analysis context where we are looking at structural health monitoring. They are in principle very very complex because you have all the possibilities in the world to collect additional information and do this and do that depending on what you find and depending on a lot of other things which may happen in the meantime. So well we need to try to find a way of describing this in a categorized manner which we can then find solutions for the categories. Now and lots of idealizations. The only last thing I wanted to say is concerns the necessity that we look into other disciplines like in the economics theory where of course they do lots of decision analysis for instance in options real options analysis. We found something when we were working for Annette and and Kassu Kassu Yoshinishima they found some very useful techniques from options analysis they call stochastic meshing in order to analyze these very very large event or decision event trees. And similar techniques may also be applicable for the type of applications we are looking at here but there could also be other things hiding in some other scientific fields so we need to be open for that. Okay I didn't take two hours I intended to take ten minutes so I guess I'm good thanks for your attention.