 I worked at the Defense Department and actually I've been working there for almost all of my life. But I started there as a Apache helicopter pilot. And later on I did my PhD in mathematics. And I always tell my students that doing math, doing game theory and graph theory and applying it within the military operational domain is much more exciting than flying Apache helicopters. And I think this is really true. The field that we are working in, the data science during the last year is becoming more and more exciting. And doing science in general and finding things out I think that's really a very great and worthwhile endeavor to undertake. So today, before I start with my main topic, I want to take you back several years. This conference is called Predict. Prediction is about the future of course. But today, first I want to start with a short story. And it is about 14 years ago. I want to take you back to April 2003. And April 2003 wasn't only the moment when the human genome project was finished. The complete human genome was mapped. April 2003 wasn't only the moment when 50 cent had a huge hit with his single in the club. But April 2003 was also a very important moment because of the U.S. invasion in Iraq, as most of you will remember. The invasion was going on several months before. But in April 2003, the U.S. forces entered Baghdad and occupied the area. The funny or maybe strange thing about this is that days before the U.S. troops got into Baghdad, there were still pictures circulating about Saddam Hussein when he was walking the streets in Baghdad and he was talking to people. He was talking to a crowd. Maybe some of you know these pictures. But in one of these pictures, you could see him standing in the crowd. And behind him, there was this very tall big guy with a brown khaki shirt with big sunglasses. And the intelligence people didn't really know this guy. They thought he was a bodyguard. And they also thought he wasn't very important. Of course, during those moments, Saddam fled. Nobody knew where he was. And it became imperative to find him. For this, the Defense Intelligence Agency and also other intelligence agencies made a deck of cards. And it is often done within counter-insurgency campaigns that you make a deck of cards with faces of important people or of the people who you think are important within the insurgency and you put them on the cards. So the highest ranking card, in this case the Ace of Spades, of course, was reserved for Saddam Hussein. And other cards were the people who were taught important were put on the cards. For instance, the Ace of Hearts and the Ace of Clubs was given to two sons of Saddam. And these cards were handed out to personnel operating in the area and especially to personnel operating around Tikrit because Tikrit was the area where Saddam was born and where they thought was the highest probability of finding him. The days went on, the months went on. Colonel Hickey, who was a U.S. Army Colonel was responsible for searching the area. They were doing raids all around Tikrit, but they were not really finding Saddam. Next to this deck of cards, they also had a blacklist of 100 more individuals. But the problem was that all these individuals were former regime members, former Baath members, members of Saddam's party, and they were rounding up these people, but they were not getting closer to Saddam. And people were getting worried. Luckily, Colonel Hickey, the U.S. Army Commander, also had a sociologist within his team, and this was Brian Reed. Brian Reed was an Army Major, and he also happened to have this degree in sociology. They were talking to each other and they were thinking about, you know, drawing paper-pencil maps about how to find Saddam. Then one day, they got a little bit lucky. Because they were supposed to do a raid somewhere, but the insurgents staged a murder attack, and they had to improvise. And during that improvisation, they went on to do a raid on another farm. And they thought the farm wasn't really important. But what they found on the farm, which belonged to a bodyguard of Saddam Hussein, was 8 million U.S. dollars. They found sniper rifles. They found night-vision goggles. They also found Julie from one of Saddam's wives. And they also found another important piece of the puzzle, but I will not tell you what it was at this moment. So you can take a couple of minutes maybe to think about it, but it was a very important piece of the puzzle. In general, however, if we go back forward in time, how these operations are conducted, and how intelligence agencies are supporting these operations, is by use of the intelligence cycle. And the intelligence cycle, I'm not talking about cognitive intelligence, by the way. So I'm talking about security intelligence. And this intelligence cycle is conducted many times, sometimes explicitly, sometimes more implicitly. But what is going on, that you start with this problem, you start with you have to find Saddam, or like the tragic events a couple of days ago in Las Vegas, an incident is going on, and you have to make sense of that incident. So you start collecting all the information you need. You start looking at the data you have, maybe you have missing data. You start tasking your resources. So you have signals intelligence, human intelligence. You apply your spies, whatever. You bring together subject matter experts, and you start analyzing the situation and formulating hypothesis. But what is very important in this regard, and which is fairly new within the intelligence community, is that these days there is an explicit step for the target modeling. So step five. And during the target modeling phase in the intelligence cycle, you take in the quantitative methodologist. So maybe you have a computer scientist, maybe you have a team of computer scientists, maybe you have mathematicians, etc., data scientists, and you take them into the equation. You present the problem you are engaged in, finding Saddam or something else, and these target modelers, they have specific, their purpose is to have specific models for the things that are going on. This can be social network analysis, can be hierarchical, hidden Markov models, it can be anything, agent-based models, you name it. These days, of course, we're moving much more towards machine learning kind of stuff. But the idea is that in this way, you integrate the more quantitative methodology with the qualitative stuff that is going on. Because in general, as you know, this is quite a large divide between these two kinds of people looking at a problem. So that's why this process was made explicit. Now, actually today, the main thing I want to talk to you about in regard to this target modeling is that I want to present two types of models that can be built within this domain of counter-terrorism and counter-insurgency. Of course, one of the problems with counter-terrorism is that after the tragic events of 9-11, it became much more interest into the structure of terrorist networks, which used to be very hierarchical, organized in the 70s and 80s, et cetera. But nowadays, we realize that this is simply not true anymore. They have a certain network topology. And there is a lot of qualitative talk about the network topology, but the target model people want to have more quantitative models that they can also bring into the equation. So one of the things I want to talk to you about is some research we did about network topology. And given time, I will also talk about another approach, and this is if you have a given data set of maybe of a terrorist organization or maybe of a larger social structure, and you want to find out who the important individuals are, how you can use some quantitative modeling also within that domain. First, the network topology. Somewhere in the years 2000, a video was found in Afghanistan in a training camp from Musab Al Suri, and he was giving a lecture to Mujahideen, and he was talking about setting up network structures. And in this regard, he gave explicit formulations about how network structures would look like. And of course, security and information processing are very important within this regard. Very simply put, you don't want everybody, you don't want everybody in the network structure to know everybody else, if somebody is picked up and he will be interrogated, he could give away everybody in the organization. On the other hand, you also don't want that nobody knows nobody else, because that makes it very hard to conduct operations, of course. So somewhere in between, there should be some kind of solution. We modeled this problem as a Nash bargaining problem, which is from game theory, and most of you will know John Nash, a very famous game theorist, he's well-known from the movie A Beautiful Mind, but he developed a axiomatic theory about Nash bargaining, and simply put, it says you have several, two or more individuals, and they are bargaining about a certain finite set of solutions. But each individual has a different utility for each solution, so how do you come up with what solutions you should pick? Well, this is what Nash actually modeled, and he did it very nice, and he came up with the Nash bargaining solution, which is simply taking the product of all the utility functions and maximizing it, and that is a very nice optimal solution for that. Well, in terrorist terms, you can think about it that two terrorists or two leaders of an organization are responsible for conducting operations. One is in charge of secrecy, the other one is in charge of information, and somehow they have to combine this. We developed a mathematical model to take this into account, and we were able to find several optimal network structures, so this gave us the opportunity to use quantitative methods to measure actually what kind of network structures you would expect in reality. Of course, this is more a deductive approach, because the problem with covert networks, with criminal and terrorist and insurgent networks, is that we don't have that much data about it, which we can use to test all kinds of hypothesis. Of course, there is data available, but it is much more sparse than the general machine learning data sets that are available to learn all kinds of things from your data. So this was one approach we took, and this can be an example of a type of model that you can use in the target modeling phase. Another example I want to give to you is key leader identification. Of course, within the social network domain, which actually somehow evolved from the graph theory domain, there is a lot of work done on centrality indices. How can we measure the centrality and prestige of a certain individual who is important, etc.? But many of these centrality formulas only look at the network structure, but in reality you will also want to take into account not only who is communicating to whom, etc., but also you want to use properties of individuals. We did research on this, we have some publications, and unfortunately we have to use data sets that are open source, of course, and we use the WTC911 data set, which is not very representative for good data sets in this kind of modeling, but if you want to present it, we'll use it. What we did, we developed a game theoretic model which actually can take into account individual properties, but also network properties, and you can very easily tune it. In reality, of course, you don't want to use one centrality measure to determine which individual is important, but you want to use multiple ones, because each measure depends on a certain context, and this context is discussed with the subject matter expert, and you would want to put all these different centrality measures next to each other, so you get different rankings, and from each ranking you learn something, and clearly what you get from this is not the truth, but given a certain context, you know which individual is important with regard to that context. Doing this, we were able to, for instance, develop a centrality measure which could very easily spot individuals that were important with regard to the WTC in connecting different flights. Well, very fast, that were two of the models that can be used by target modeling in the intelligence cycle. Of course, there are many, many more possibilities, but going back to the Red Dawn operation, finding Saddam Hussein going 14 years back, the way they did it was that during this raid, they also found a photo album, and this photo album was about the marriage of Saddam Hussein, and in this photo album, they gained a lot more information about the social network, simply because several individuals were seen on the photographs that were thought totally not important, but they happened to be important because when they started to put these figures also in the social network, and they did very basic analysis on the social network, they were able to find who are called the Muslim brothers, and they were the ones that actually led them to find Saddam Hussein in the end. So, closing my talk, you can see that quantitative modeling can do a lot of good things within the intelligence domain. However, it is a challenge to bring the different groups together, but I think, especially now with the coming machine learning and everything that is going on within the artificial intelligence world, I hope this will make a difference and make the world just a little bit safer. Thank you for your attention.