 Thank you very much for inviting me to this conference. So I'm going to talk about the emergent collective dynamics in large-agent systems. It's so large, so it's not fitting a screen. So my main field is artificial life. And I was studied in 1987 in Santa Fe. I was in Santa Fe, too, in 1989. So then there's tons of interesting concepts and ideas. And then still, after 30 years, we are still working in this field. And then the aim of this artificial life field is to answer the question, what is life? By exploring emergent dynamics and open-ended evolution. So this talk, I'm going to update what's the recent development in artificial life on how the artificial life is committing to biological life, I hope. So this is my drawing when I gave a keynote at the artificial life conference in 2008. So artificial life is all about the theory of evolution on the brain. And back in 1950 to 1960, there is a cybernetics. It's about an autonomous robot and then a philosophy and Alan Turing's chemical reactions, and then also von Neumann's cell automata. These are the classic artificial life studies. And then from that, we have been developing all these interesting ideas. And artificial life started back in, as I said, it's 1990. There's a bunch of interesting ideas that came out. And then using that one, we try to understand what is life and then what the evolution is. So this year, 2018, I'm organizing this artificial life conference in Tokyo. And then, so this is our home page. And then theme is beyond AI. So there are a bunch of people interested in how to use AI. But I'm sure that the era of AI popularity is going to end within five years. And after AI paradigm, we have to think about new epistemology of artificial life and complex systems. So my principle is life must come first. And artificial intelligence is just a side effect of artificial life. So artificial life is going to be the main theme within five years, I hope. Then as you know, one of the big challenges for artificial life is how to use new technologies, which we call exponential technologies. So the computer is getting faster and faster. That's the main engine why we can use deep learning and we can use a bunch of data flows. So using exponential technology, the old ideas of artificial life models back in 1970 and 1980 can be updated. For example, like the Game of Life, random VR network, and also the Red Cray Reynolds Boyd model, it also can be updated with the exponential technologies. So for the first, my talk is about Boyd model, but Boyd model not in my computer, but with a K-computer and then also GPGPU. At the Boyd model, you may know that everybody been talking about Boyd model, which is separation, alignment, and cohesion. There are a bunch of ways to write down into the equations, but I didn't check carefully, but this one is more or less like a generic forms of Reynolds model. So simply using this one that I can... So this is the famous Boyd model with 256 individuals, and you can see there's a sort of self-organization of structures and patterns is evolving from random conditions. Even though this is quite a simple toy model, but it took a lot of our computer power when back in 1990. But now we have a strong computer, so I was hoping that we can compute more complex ones. So in order to assimilate a large-scale Boyd model, we need GPGPU and then also a K-computer. The K-computer is a supercomputer in Japan, but still it's compatible with the GPGPU because the K-computer is even fast that many people want to use it at the same time, so it becomes slower and slower. So this is the periodic boundary condition in a three-dimensional space. Then the numbers of the individual is 2048, and so there's a big flock is going from here to here. So going from the left to right, that's what you can see here. Because then the next scale is 16482 to the 11th. Then this different kind of structure emerges, like this one, which is connected by this small snake-like structure. The next scale is 130,000. By the way, the density is the same. The space is getting bigger and bigger to increase the numbers of the individuals, but the density plus space is always kept constant. And then you can see some interesting structures in there. Then you can even scale up to the next level. So this is a half million numbers of the birds. It's not birds, but it's individuals of making flocking behavior. So this structure is something interesting. When I gave this talk, the biologist studying bird flocking behavior raised his hand and said, he said this is not birds, but I knew that. So this is something like abstract flocking behavior that you can simulate with a bunch of birds. So one of the big questions with this system is how to identify each flocking behavior. So usually the people assume there's one simple flocking behavior, but actually what happens is there are many, many of them. So I use K-means, it's one of the simplest clustering methods, a self-organizing map, or a DB scan, or even deep neural network to classify what is flock and how many flocking behavior, flocking patterns that you can find in this space. That's what I was working with. Then one of the methods that I was using is not three-dimensional real space, but we have to use four-dimensional metric space. So XYZ plus the amplitude of... Sorry, it's my Gmail. Amplitude of the speed is also the important dimension. So this is the one that you see. So when you see this, for human observers it's apparent that there is a flocking behavior which is connected by the snake-like patterns, but if you want to use machine learning to classify them, usually it's difficult. I mean, if you are scaling up, there are a bunch of interesting behaviors happening. I will show you that within this flocking behavior there's a brownian or a singular levy-flight-like behavior you can find in here, but outside of this one there's a more coherent dynamics that you can observe. Then... So by using some techniques that we can temporarily follow how the flocking behavior changes over time. So you can see some of those flocking behaviors, like blue ones and all these kind of things, it's changing its patterns, but these are very slow dynamics. It's slowly changing its structures, but for these snake-like patterns, they have less lifetime so that they can vanish, but they are very interesting roles to connect from one big one to the other big ones. So for one of the techniques that we can use is non-negative matrix factorization. So taking... Usually the people use it as a document in terms, but for this void model I use... This one is the time axis, and this is the ID number of the individuals. So we can factorize into some sort of mode. Maybe these are corresponding to the flock, and then this is the IDs of the void. So this is one way to classify whether there exists flocking behavior in terms of NMF. And you can see... This is time from 1,500 time steps to 250 time steps. There's some waves coming up, and these waves are corresponding to these patterns, and this one is a kind of mode, so there's one. So you can see some of the mode. So even if you have a big brogue, and there's a bunch of different modes is coupling with each other and then evolving in time. Then we look into the speed. So this is the velocity, and this is the fluctuation of the velocity, and this is the density, local density, and then also this is the local density fluctuations. So if you look at the size, then the bigger, the larger size swarms have a slower time scale, slower speed, and then the density goes up. So... And then also the fluctuation has the same thing. Small to middle-sized flocks have larger fluctuations, but for the larger flocks have larger local density fluctuations. So the big one has a slow speed, and then fluctuation goes up with the density. But the middle-sized and then also small ones has much faster than the big ones, but the local density fluctuation is very suppressed. And as I said, there's an enormous diffusion that's going on in the large flocks. So in sum, what happens is there's two different kinds of fluctuations that exist. This is the floc size, and this is the susceptibility. So when you see... So the floc size is going from the scale of 10 to the second to the 10 to the fifth, and the bigger ones have fluctuations over density fluctuations, but the smaller ones, less than 10,000, or 1,000, is about velocity fluctuations. So the different susceptibility to the different external perturbation can be expected from here. So that's what I thought was quite interesting, because when you are scaling up up to larger than 10 to the fourth, then you can find a different kind of structure which is characterized by these different kinds of fluctuations. That's what I thought was interesting with these large systems. And also, people might say, well, because this one is very abstract model, but I have been working with this oil droplet, which is oreic anhydride with water. Then this is my very first experiment I did in 2007. So you can see a bunch of droplet is generated and then coupling with each other and moving around. So this is more like chemical gliders. The game of life is very fragile, but this one is much more robust, because this one is happening in the real space. So the one that I saw in the Boyd model can be applicable to this kind of artificial chemical droplet. Then I was thinking that maybe the one that we have found can be applicable to the honeybee stuff. So fortunately, I've been working with Jean Robinson in the University of Illinois. Then we have analyzed how the honeybee hive is going to show any kinds of corrective behavior or not. So we have five different honeybee hives. Then each bee has a QR code on its back. So we can trace each individual for a longer period of time. Then we can see what sort of individuality can commit to the corrective behaviors. Then I can tell you one of those discoveries that I found. First of all, this is a video that we can generate from the data. So this is where the heading of each bee, and the data is like this one. So this is time, and this is the positions in two-dimensional space, and this is the heading direction. And this is the bee ID number. So when the line is connecting to one individual to the others, it means that the bees are interacting with each other face to face. I don't know whether you can notice. Sometimes there's a bunch of, you know, this collective activity is going on, and then it's calmed down, but it's coming up. There's some collective bursting behavior that can be generated within the system. So now it's bursting, right? But now it's going to suppress. Again, it's going to be happening again, right? So in order to analyze this one, we look into this time series, and this is kinetic energy. So we can define kinetic energy for each individual bee, like for each unit of time, how much distance each bee can move around. So this is the definition of kinetic energy. So we can define total kinetic energy of the nest. So this global kinetic energy is bursting over time, and here this door was open. So first of all, the honey bee hive was in a closed set, but then the door was open, so honey bee can go up, go outside, and then come back. So the question is how this bursting behavior changes when we open the door, and then before and after the door opening event. Then we noticed that there is an interesting two different characteristics of bursting behavior. One is an endogenous burst, which might be caused by the interaction between bees, and there's an exogenous burst, which is caused by externally. There's kind of stimulus is coming from outside. That's why the bee becomes activated, and then there's big burst is coming out. So I look into this. So this is time, and this is how the global kinetic burst is going up. And you can see this is for the endogenous, and this is for the exogenous. So endogenous starts very slowly, but exogenous happens rapidly. So we try to understand what causes by this kind of thing. So because we have a QR code for every bee, so as you see here, right, so this is the global kinetic energy. So before this global bursting happens, first of all this B574 is activated, and then another one is also activated. But this one is a queen bee, so it's always activating, but this B574 or B31 is only happens at a certain period of time. So in this case, B574 and B31 started to burst, then the other one is also start to burst. So we try to, again, we are using like NMF, regenerative matrix factorization. See what sort of a mode we can see here. So for this exogenous burst, there is only one or two, I mean, it's maximum, almost all the bees are contributing to this bursting behavior, but for the endogenous ones, there is a sequential activation of bursting behaviors. First of all, some bees become activated, and then those bees are going to activate some other bees, then other bees are going to activate. So there is a cascading processes going on here, but for the external exogenous burst, there is no such cascading bursting behaviors. It's more like everyone is going to activate at the same time. So we are using, this is a multi-dimensional scale. So which one causes the burst? You can see, so this is the different nest. This is 1201 and 1301 is a different conditions with the different numbers of the bees, but still the red one happens before the door opens, then the blue one happens after the door opens. So before the door opens, the bees are only in the closed room, but after the door has been opened, the bees can go out and come in. So there is an information cascade that is expected after the door has been opened. And then also interestingly, the individual bees that cause the burst is different before and after the door opening. So maybe some of the bees can carry the information and then causing an information cascade when the burst has been observed. So we try to do whether the foragers can be a pioneer bees. So when I say pioneer bees, that can cause bursting behaviors. And as you see here, once you open the door, there are a bunch of bees which go out to forager, to become forager, is also committing to the burst. So maybe two, the foragers can be the pioneer bees. That pioneer bees can cause bursting behaviors. So there should be some sort of information cascade is going on when you see these corrective behaviors. And then finally, one of the things that I want to discuss is the web services. So the first was a void model that we can see some collective behaviors of the big flocks and then was connected through this coherent behaviors. But there's two different fluctuations that we cannot observe. And then honey beehive, there's also different kind of clusters which maybe information is carried out from the foragers to cause bursting behavior. And those bursting behaviors is a characteristic of information cascade within a honey beehive. And web services is interesting because it's one of the most complex artificial systems that the human being has been created. And then there's one of the services called RoomClip that we are using RoomClip data to analyze what kind of evolutionary processes is going on within web services. So this web service is a tagging data. So each user can post a photo with some tags associated with the photo. So we have each user's ID and also what kind of photo or what kind of tag has been submitted to the system. And then the system size is about 10 to the... So it's almost 10,000 to 100,000 users has been involved in this web services. So what is interesting here is that... So we are using, like, Hawke's processes. Hawke's processes is like Poisson processes. So this one is Poisson processes without the second term. But with the second term it's more like a feedback from the other users. So once some user posts the photo then other user posts other contribution to the web services. If there are some interactions, if someone posts then the other one also wants to post again and again. So there's a feedback from the other users. That is represented by the second term. If you take into account this term then the interesting point is that... So the integral kernel is an exponential term. And once exponential term goes to 1 then it's going to be critical. So what I'm saying here is that if you can take an integral kernel as A times exponential to the minus B then this A over B is a nice index that you can see how the system is evolving towards what. So what you can see here is that the exponent in A over B is going to close to 1 as time goes by. So this is the time and this is the exponent. So exponent is starting from here and then going towards 1. But not exactly 1, right? So what we can see is that the exponent is somehow distributed around 1 means that the system is going to the critical state. You can say this one is like a self-organized criticality but it's not on the critical point. It's a little bit before the critical point that the web services are diverging into this point concentrated on this point. And more interestingly, we are looking at what kind of users is committing to these web services. So we are looking at the active users and then using a clustering algorithm to simulate, to characterize what kind of users that you can see here. And then distance between users is measured by the Jensen Shannon divergence, which is if the two users are using the same keywords, same tags, then we say these users are close to each other, right? So at one threshold, we see three different communities. Three different communities is spontaneously emerging from these web services. And these three different clusters are corresponding to they are the same users. This red one and blue one and yellow one, they consist of rather similar users. They have similar profiles. They have been using similar keywords. That's what I thought was interesting. And then more interestingly, when we are looking into who is generating the novel keywords, right? So RoomCrip users can generate, sometimes put new keywords, but sometimes just reusing the already existing keywords. So this one, the novelty is coming from same users. That's what I thought was interesting, that we're not going into the details, but if the same user is collecting to each other, then those clusters with the same profile users tend to generate more novel keywords rather than heterogeneous communities. So I can say that novelty or creativity is coming from, not from a heterogeneous community, but it's a homogeneous community has more potential to create new keywords. So my discussion is that... So starting this large system, also the evolutionary system that you can discuss that... Well, you can develop superorganisms, but you really have to increase the numbers of the size up to, say, 10 to the fourth. So 10,000 is probably the first critical size of the system that can create superorganisms. In our cases, different kind of fluctuation emerges after 10 to the fourth. So what we can expect is after this one, it's probably one billion numbers of those sizes. There's a new kind of superorganism that can emerge. So like in our physical systems, probably there are orders of the scale orders that where the superorganisms can emerge is determined by this number. But this is my hypothesis, but I don't know whether that's true or not. The second thing is, nobility is coming from homogeneous groups and also isolated groups, and they can make nobilities. So usually the people think that community must have heterogeneity to have interesting diverse cultures and new culture to come up. But actually what happens is that homogeneous groups can create more nobility comparing with heterogeneous groups. And the third thing, I didn't have time to introduce here, but robustness is also possible when you have a huge number of individuals. So what I did was, one is a Peter Gatchis model. So this cellular automata has two to the 223 states per site. So it's more like a cellular automata with each site has a big computer. Big computer is connected to each other. So if you have this system, it's self-simulating what happens in your own system. So using self-simulation that you can be much robust to the noise. This is also the von Neumann's idea that, von Neumann's dream, how you can make a self-reproduction which is robust against noise. And then what is required here is not just two states or three states, but two to the 223 states is necessary per site to have a robust cellular automata. That's what we can expect. And we, 70% have done with the simulations of the Peter Gatchis model. But maybe next time I can show you. And finally, my final quote is that this is a paper from Rodney Brooks, which I also invited him to the artificial life conference this year. And he, in 2001 Nature paper, that he says why his robot or why artificial life never becomes life. That's a huge problem for artificial life people. Why the Roomba is still robot and not the living systems. And then he discusses within his paper that maybe the first one is we might be just getting a few parameters wrong. Or maybe we might be building models that are below some complexity threshold. Or maybe perhaps it is still a lack of computing power. And then I thought maybe we have to go to number two and three. That's why I'm always making complex models. And then also because we are now having a big computing power, so two and three might be solved. And I hope my selection of parameters is right. But always people want to understand what's missing. I can be missing some fundamental currently unimagined laws that may exist in the world. And that's what I still dream sometimes. Maybe not probabilistic or not deterministic, but should be in deterministic model. Should be applied to understand what is life, to answer what is life. And that's the future of artificial life studies. Thank you very much.