21600 >> Yuval Peres: We're happy to have Elliot Anshelevich from RPI. His topic is very topical these days and a lot of interest here. So I'm glad that the talk will be recorded. So we can look more at these problems, because I think these problems are going to be central for a long time. Please, Elliot. >> Elliot Anshelevich: Okay. So this is going to be a talk about games and social networks. All right. So the idea is let's say we have some social network, we know there are people, whatever, and the links between the people represent relationships or collaborations or friendships, something like that. Okay. So I mention there's a big social network, which I didn't draw, and we're this green person. And these are my friends. So I just drew the links that are next to the green person. Okay. When the green person like all of us has a problem, that problem is that you can't possibly put 100 percent effort into every single thing you're working on or 100 percent into every friendship that you're participating in. So we're going to model this by saying, okay, this green person has one unit of effort. And they're going to allocate this one unit of effort amongst all their relationships that they're involved in. So, for example, it might do something like this. So here are maybe this green person puts 60 percent of its effort on this relationship or this collaboration. It's working really hard trying to make this relationship succeed. And maybe this person puts 20 percent on this one and this one. 20 percent is actually a lot. Put some small amount on those two. So it means that the green person in this relationship or this collaboration is sort of putting in just enough effort to push it forward, not really working that hard but just putting in enough to make sure it keeps going. And then these two friendships the green person is completely ignoring, letting them take their own course and putting in no time at all. So this is a pretty common problem that occurs to people. And, okay, so what will happen if let's say that this is a particular link. So this is a relationship between two people, orange and green. And what happens if they both put in a lot of effort into this friendship. Well, presumably what should happen is this relationship will succeed. If it's a collaboration, publish lots of papers or something, or if it's a friendship, then they'll both be happy and they'll both become good friends and get a lot out of this friendship. So they'll both be happy because of it. What happens if neither person puts in very much effort? Well, as unfortunately a lot of us are familiar, if neither person works very hard at a relationship or collaboration, it will just die. Eventually, maybe not for a while, but usually it will fizzle out, nothing will happen if nobody's trying, neither person will be very happy with it. Finally, what if one person puts in a lot of effort, in this case the green person, and the other person doesn't put in much at all, the answer depends, it depends on the kind of project or kind of relationship or friendship that it is. Sometimes, I mean, it's possible that actually they'll both still be quite happy because of this. It's possible that this collaboration will succeed, even though this person is not putting in much effort at all. They'll both reap rewards and they'll both be pretty happy because of this. So it sort of depends. We're going to try to model this as generally as possible or at least I'll try to be as general as possible. So I'm going to say that -- okay, let's say that this edge is, let's call it E. So for every edge E in the graph we're going to have some reward function. So this function F sub E. It can be different reward function for every edge. And it's going to take an X and Y. So X and Y are the efforts that the two people are putting into this project or into this relationship. So this function takes X and Y and spits out how successful this friendship is. Spits out how the success or the number of papers or whatever, how happy these people are with this friendship. So it's an arbitrary function of two variables. But we're not going to try to not assume too much about it. We're going to say this function is nonnegative. And it's nondecreasing in both variables. If you put in more effort, you should only get a better relationship or better collaboration. Not worse, right? I hope it's true. >>: Symmetric X and Y. >> Elliot Anshelevich: Doesn't have to be symmetric. >>: You're assuming [inaudible]. >> Elliot Anshelevich: I was just about to get to that. This function doesn't do symmetric X and Y. What happens is both efforts go into a function. The success of the project comes out and both people get that benefit. So I'm assuming exactly that the success of a project or benefit of a project is the same for both participants. You can imagine, in fact maybe I'll talk about this a little bit at the end of the talk, you can imagine if somehow there's a total amount of benefit and they're dividing it amongst themselves or get to somehow unequally -- that's not what this talk is going to be about. There's some extensions we can talk about but that's not what this talk is about. In this talk you think of it like they published some paper together and both people get the credit. And if the paper is good, they both get the credit. Or you can think of it as they're doing this project together and their boss sees the outcome of the project. The boss sees you did well and you both get promoted or something. The boss doesn't realize that this guy didn't do almost anything. But that's how it is sometimes. So for this talk the assumption is they both get the same benefit from the edge. Okay? Okay. So this is sort of the introduction. This is a setup. Now let's actually talk about what is the actual game. Define the game that I'm going to talk about. So the game is as follows. So we're given an undirected graph. This is the social network. We know exactly what it is for this kind of talk. So we're given the entire graph. The players in the game are going to be just the nodes of the graph. So every node is a player. And so, for example, maybe here's a very simple graph with three players. Right? And each player has a budget of effort. This is the budget they're going to allocate amongst all the edges that participate in. Now the budget can be different. So here we have budget of 5, 10 and 8. Why is the budget for different players different? I think it's very natural. I can make them all the same. Some nicer things happen if they're all the same, but I think it makes perfect sense to be different I certainly know people have different levels of energy. People can put in more effort than others in total. If you talk about something like friendships, I certainly know people who spend more time on their social lives than others. I think it's better to have their effort levels to be very different and we're given the reward functions. In this case the reward functions are very simple, 3 times X plus Y or 2 times X plus Y. In general they can be arbitrary reward functions. This is what we're given. And what are the strategies of the players? It's exactly what I said. Each player is going to take their budget of effort and divide it amongst the friendship it participates in. So to put this in some notation, although I don't think we'll ever use this notation in this talk, but just to make it formal. So each node is going to, okay, this node, for example, takes us ten years of effort, say I'm going to put in four here and six here. And, in fact, you're never going to really keep some effort to yourself. You're going to use it all, right, because the functions are nondecreasing. So there's no reason to keep it to yourself. And to formally write or contribute some amount to each edge, total amount it contributes at most to the budget. Okay. So is there a question? >>: So there's no benefit to just relaxing? >> Elliot Anshelevich: In this model, no. That's a very reasonable extension to have, actually, which I'll try to mention at the end. Yes. But in this case, yes, it's a budget constraint as opposed to an elastic constraint. Here it's just a total budget and you want to spend it all. Okay. So this is what happens. So people choose how to allocate their effort. And now that they allocate their effort we know what their reward is going to be for every edge. 15 is just three times four plus one. So now we know what the rewards are going to be for every edge. And then what's going to be the utility of a person? Well, it's just going to be the total reward they gather from their edges. So to put it another way, this person's going to get 15 plus 22 and to put it into formal notation, the utility node is going to be the sum over all the edges sent to them of reward in those edges. So that's all that means. Okay. So that's the game. So notice that the utility, the reward of a particular player depends, okay, certainly depends on how they allocate their effort. It also depends on how their neighbors allocate their effort. But it also sort of indirectly depends on neighbors of neighbors and so on. For example, you can imagine some nodes over here, this node over here decides to put a lot of effort on the edge with this node. Because I put a lot of effort over here, this node labeled eight might say I'm going to put a lot of effort over here too. Depends on the reward function. Now that I put a lot of effort here means I'm subtracting the effort here and now this person is worse off. Have this sort of behavior where what one person does affects a neighbor, which affects a neighbor and propagates and various dynamics and so on. And another thing to know since we're looking at this picture is, well, the reward functions are different, right? So here even though because the reward function here has four and here has two, even though people are putting in less effort, you have 28 instead of 22. That certainly makes sense. Reward functions should be very different, because what do reward functions represent? One of the things they represent is compatibility of the pair of people. For example, if two people are very compatible, then a little bit of effort might give them a lot of reward. If they're not very compatible, a lot of effort won't give them very much reward here. Makes sense the reward is high even though the effort is less. This is the game. Any questions? So that's the game. Now, what are we going to look at in this game? I'm going to look at very stable solutions, think about their quality and how to find them where they exist and things like that. So I'm going to look at the equilibria. We'll look at natural equilibria. But I'm going to try to convince you that natural equilibria is not very interesting here. What do we know about natural equilibria in this game. We know pure natural equilibria exists. We know there's a potential function that matches best responses, which we means we know people always converge to natural equilibria, by doing best responses. We even know that the optimal solution, there's the one of maximal social welfare, is a natural equilibrium. So we know a lot of stuff. But they're not very interesting or they're not the right concept to use here. To give you an example. Looking at this example, let's look at this edge. Imagine I have a thousand, some big number, times X plus, X times Y. So if that's the reward function here, then really what should happen is that both of these people should put some effort into this edge, they'll get a huge reward. But in the solution I drew here, this person is not going to move any effort into this edge, right? Because for this person's point of view, if I move some effort into this edge, how much reward am I going to get? Zero. Because the other person didn't. >>: This shows there's some bad equilibrium. >> Elliot Anshelevich: Shows there's some bad equilibrium. >>: But the best. >> Elliot Anshelevich: The best one is as good as social optimum. So if you want to say, okay, imagine the single edge, right, where both people are contributing nothing to this edge where the function is something like this. Then natural equilibrium will contribute everything. That's natural equilibrium, social optimum. But to contribute nothing is also equilibrium because you need both of them to do things together. >>: But this is something that often occurs. >> Elliot Anshelevich: This is something that often occurs. Something that often occurs. Some equilibria are really bad and some very good. >>: Naturally will occur, I think, in practice, because ->> Elliot Anshelevich: That's the question. >>: They'll be great collaborators without knowing each other. >> Elliot Anshelevich: This is the way the links exist, right? >>: Or they don't know each other. >> Elliot Anshelevich: If they don't know each other, this link shouldn't exist. So feel free to disagree. My view here is that I think for this kind of game, natural equilibria don't quite make sense because in real life, whatever that means, I think these two people who know each other, because there's a link between them, would talk to each other and say if we both increase our effort a little bit, we'd both get a huge reward. So I think they would do that. Or to put it another way I'm going to look at coalitionnal equilibria. And if you like natural equilibria in this, game I just told you all the results that I know. So quite a lot of things are known about natural equilibria as well. But here's what I'm going to do in this talk. I'm not going to look at natural equilibria. I'm going to look at a more interesting concept for this game, and that's pairwise equilibrium. What is pairwise equilibrium? Okay. So just to define some pretty usual terms. Unilateral improving move means that a single player, a single person, can change their effort to location and improve their utility. A bilateral improving move means that a pair of players can both change their strategies, can both change their effort and locations at the same time, and both of them strictly improve their utility. Okay. So natural equilibrium is exactly something that has no unilateral improving moves. No single [inaudible] but a pairwise equilibrium is something where, which has no unilateral or bilateral improving move. Every pairwise equilibrium is also a natural equilibrium but not vice versa. And, in fact, I'm going to -- so most of the results that I'm going to tell you actually are going to hold for strong equilibrium as well. A strong equilibrium is even the stronger concept than pairwise equilibrium. In fact, that's why it's called strong. It's a really strong concept. Means no -- equilibrium means no pair will deviate, no coalition, no set of players including all the set of players will deviate such that every single person in the set will improve. >>: Core. >> Elliot Anshelevich: What is a core? Core is you have a coalition and they pool their money, transferrable utility. And they deviate such that the total utility goes up. Strong equilibrium means that it's nontransferrable utility. So each person has to improve. So strong equilibrium is the equivalence of the core for NTU games, for nontransferrable utility games. Okay? And the games I'm talking about are nontransfersable. It means that every person has to improve. It's not somehow they get together and pool their money. I guess it's not so easy to do with friendships or with credit from papers. You can't quite say like if that -- oh, since you're better off and I'm worse off I wouldn't necessarily be happy. You want that everybody improves. Any more questions about transferrable utility and nontransferrable utility? So in this talk I'm going to look at pairwise equilibria for this game where you partition your effort amongst your neighbors. And I'm going to ask these kind of questions. Okay. First of all, I'm going to ask: Do pairwise equilibria even exist here? So as I mentioned pure natural equilibria always exists. That's easy to show. But a pairwise equilibrium as you see doesn't always exist but I'll show you when it does. I'll ask about questions about price of anarchy for this talk I will mean the relation of the social welfare in the optimal solutions, maximum social welfare possible. And the worst pairwise equilibrium. So that's what I mean by, that's what I mean by the price. Price stability, I can also give balances for optimal solutions versus the best pairwise equilibrium. You can ask me afterwards if you want. >>: Compute, you're going to add up all these nontransferrable things. >> Elliot Anshelevich: Right. Social welfare. Social welfare is the total utility of everybody together. Which is in essentially the friction between total happiness and happiness of particular people. Because social welfare might go higher but a particular person might not like it which is why they would deviate. So yes. But I think I can see your point also. So price of anarchy, we'll look at price of computation, pairwise continuum, when it exists, we can compute it. And we're going to look at basically can the players compute it, meaning are there some nice reasonable dynamics where players just take turns deviating or something which will actually converge to pairwise equilibrium, when they exist, of course. So these are the kinds of questions I'm going to look at in this talk. Let me mention very quickly, let me mention related work. First of all, there's a lot. And many more. Social networks in general and games and networks, all this is a huge area. There's lots of stuff that's related. Ask me afterwards if you want to know more. But let me mention a few of these things. All right. So even things like stable matching. So utility maximizing version of stable matching is actually an integral version of a very special case of our game. So there's some relationship there which you'll actually see probably towards the end of the talk. Network creation games. There's a lot of network creation games of all kinds. Probably the most relevant network creation game to this is the co-author model from 1996. There's lots of differences in this model you don't spread effort among edges. You sort of choose -- you construct edges. So you form edges there or not as opposed to having some effort on the edge. There they have a complete graph instead of some given graph. They have very specific reward functions, notion of pairwise equilibrium. Lots of differences, but I think this is the most relevant version. And there they also look at pairwise equilibrium because they want people to be able to form edges in pairs as opposed to sort of one person forming an edge and then later saying okay I'll do it too. Okay. So the game I mentioned, this contribution game, is technically atomic splittable congestion games, for those who know congestion games. Congestion games are a a very big field. Lots of results known. Atomic splittable congestion games are sort of the most annoying ones to handle. And there's lots of differences between what we're doing and various congestion game results. For example, usually in congestion games people care about natural equilibria and minimizing costs. We're looking at pairwise equilibrium and maximizing utility versus maximizing utility different when you look at price of anarchy, if you look at optimal solutions, it's the same. >>: How are you splitting the atom here? >> Elliot Anshelevich: So atomic splittable -- I didn't think of this term right. Atomic splittable means that each person controls some big amount of flow, in this case effort. And they can split it amongst different parts of a network. >>: Why is it atomic? >> Elliot Anshelevich: It's called atomic to differentiate from nonatomic. Nonatomic means that each person controls an infinitesimal amount of flow, something that approaches zero. Therefore nonatomic. I didn't make up the name, but yeah. I mean, there's many other differences. Like, for example, usually the delayed function, congestion games, correspond to reward functions. In this game are some function of X plus Y. In our case they're arbitrary functions of two parameters. Okay. I should mention in some public goods and contribution games because they had the words contribution games in them. These are very different games. Although, I think they're very interesting and there's lots of cool stuff in there but they're very different. Here's how. Those games, each person has one number as their strategy. So you just decide how much you want to contribute to society as a whole. So you don't like say how much do I want to contribute to each project or each relationship. We just figure out how much they want to contribute to society. And that's your strategy. Depending on how many people contribute, various nice things happen. But I think it's actually very interesting area. Finally, last thing I'll mention is minimum effort coordination games. So one of the reasons we started working on this topic is because of minimum effort coordination games. These are in fact a bona fide special case of the games I'm going to be talking about. Studied in various economic papers. Usually experimental. So nothing was proven about them until I guess this work, or almost nothing. But some very interesting experiments were done. Some on actual people as opposed to simulation, and I think they're pretty cool. What are minimum effort coordination games? It essentially would be a game I mentioned, but the reward functions are the form min of XY. It's the minimum of a contribution that matters and that's all. So you can see how they'd be much more structured because of that and many more. So this is all I'm going to say about related work at least today. So let me tell you then a little bit about the results that -- what do we know about these contribution games, basically? So what I want to do is I want to talk about various properties of these games, depending on the type of reward functions that are present. So as you probably can see, what the reward functions are like can make a huge difference. For example, if the reward functions are convex it kind of means I want to put more of my effort on the same edge. If they're concave you want to spread out your effort. So what I'm going to do I'm going to classify various properties, specifically existence of pairwise equilibrium and price of anarchy. Depending on the type of reward functions present in the network. So, for example, suppose that all the reward functions are convex. They don't have to be the same function. They can be different for every edge but suppose they're all convex. Okay. Then pairwise equilibrium always exists with a caveat. So pairwise equilibrium always exists if both people have to contribute a little bit of effort to get non-zero reward to every edge. If this is not true, then it may not exist. In fact, it's NP hard to figure out given a graph if it exists or not. So that's existence. And I'll tell you a little bit about more about this result in a moment. Let me just go into price of anarchy. Price of anarchy is always 2. Pairwise of equilibria are factor of two with optimization. Except there's a caveat again which I'll tell you about in a moment. But this is only true if a reward function all exhibit digit complements. I'll tell you what this means on the next slide. Annoying. I don't like these caveats, but that's the truth. So that's how it is. Okay. What if all the reward functions are concave? No caveats here. So here it may not exist. They may not be any pairwise equilibrium at all. But if it does exist, it's within a factor of 2 of the optimum solution always. The next thing is I like this special case. It's sort of a very nice structure special case. This is when all the reward functions are X plus Y times a constant. And the constant can be different for every edge. But other than that, all the reward functions are the same. >>: Why would you put all of it in this one edge? >> Elliot Anshelevich: In this one? Yeah. In this one you would -- if the Cs are the same, you would split it. This is actually a very easy case. Here's what happens, right? So, first of all, price of anarchy is one, that's obvious. Each person is always independent from the other person. Just put it all in the edges of best C. The only reason I'm bringing this up here is I think this is interesting. So pairwise equilibrium may not exist given a graph. Natural equilibrium always exists, remember, but pairwise may not. I can convince you why, but maybe afterwards. Basic thing is the pair is deviating. So even if they're getting a very high reward, they put it on the maximum CE, they might still want to move their effort to each other and the reward will increase by twice this, because I'm putting stuff here and they're putting stuff here. And the reward goes to both. That's my two-second version of why. If you didn't get that, that's okay. So it may not exist. But figuring out if it exists or not is efficiently solvable. And I think it's kind of a cute algorithm for getting out. That's why I put it up here. Okay. Let's go on. Minimum effort games. Number of minimum effort games, what I mean by this is all the reward functions are something like convex function of min of XY. Then, okay, this is the last caveat, I promise. So here, again pairwise equilibrium exists, price of anarchy is two if all the budgets are uniform, if they're all the same. It would be NP hard otherwise. Notice this is not a special case of this. This 2 doesn't come from here. And this doesn't come from here. So convex of min is not convex. So this is not a special case of this. Minimum effort concave is concave of min. Okay. Concave of min is concave. So this just comes directly from here. So that part's easy. But now we have this. So the fact that nonconcave of min gives us the very nice statement which I'll tell you about in a few slides that now pairwise equilibrium always exists when it didn't before. Maximum effort frankly is boring. I'll just include this for completeness. It's not very interesting. It's not very hard. I think this part is more interesting. If you want to look at approximate equilibrium, the optimal solution, the one of maximum social welfare, is two approximate equilibrium, two approximate pairwise equilibrium. Meaning no pair can switch and both increase their utility of more than a factor of two. That's it. So that solution is nice. So as a result, there's other things which I won't mention like price of anarchy bounds are tight. There's price stability results. But I promised I would tell you something about convergence. And I'll tell you very quickly the idea and try to give you the details if I have time. Okay. So here's what we know about convergence. Okay. Here there's two checkmarks. What it means is things are really nice. In this case basically whatever deviations or whatever dynamics you choose, they're going to converge to pairwise equilibrium if it exists, of course, and they're going to do it really fast like N squared time. Here there's only one checkmark. Things are not as nice. That means that things will converge. But they might take a very long time they get to pairwise equilibrium. But pairwise equilibrium always exists so they'll get there eventually. Here things are sort of weird. The answer is kind of yes and no and I'll tell you about that at the end, because I think it's sort of an interesting case. So the remaining time, what I want to do then is give you an idea of why some of these things are true. I'm going to start with both this red box. So really there are two theorems here. The first one is if the reward functions are convex for all edges, and if this thing holds, meaning that for every edge both people have to contribute at least a little bit to get positive reward, so I guess it should say FE of X0 equals FE of 0Y equals 0. If you don't contribute anything, if one of the people doesn't contribute anything then there's no reward. Pairwise equilibrium exists. Otherwise it may not. And in fact it's NP hard to determine if it does. The second theorem is if all the functions are convex and exhibits digit complement then the price of anarchy is two. What's digit complement? Let me tell you the words first. It's going to mean this but why this. Digit complement is a pretty standard term in economics. Probably most of you have heard it before. Here's what it actually means. In words it means if I put in more effort, then your incentive to put in more effort only goes up. The more I do the more you want to do. The strategic complement. Okay. So what does it mean for how do we write that for a function? It's just this. So the second is nonnegative. So the digit complement is opposite, substitute, where if I put in more effort your incentive to put in more effort goes down. So this is digit complement. This is some of the things -- do you have a question? So X increases the derivative of Y. >>: So why, some version of ->> Elliot Anshelevich: Yes. It's a version of -- this goes up, right? >>: Similarity? >> Elliot Anshelevich: Unfortunately, convex doesn't mean what you want it to mean. Really if I had my choice, maybe this is what convex would mean, because it means like what -- okay. So follows digit complement. For example, these functions all exhibit digit complements. Any polynomial of positive coefficients is going to exhibit digit complements. For all of these, this theorem will hold. And for these two, they also satisfy this thing which means it will also pairwise equilibrium will also exist. But not for these two, but in fact I can give -- if you look at a triangle with reward function being 2 to the X plus Y pairwise equilibrium doesn't exist. But when it does, price of anarchy is two. So let me give you an idea why these theorems hold. First one is quite easy. Here's an algorithm for the first theorem which creates a pairwise equilibrium. Always. Take your edges. Sort them from highest weight to lowest weight, where by weight, I just mean, okay, if both participants put all their effort on this edge, like all their budget on this edge, then that's their reward of this edge. So say that's the weight. Certain highest weight to lowest weight and now do a greedy matching. That's it. That's a pairwise equilibrium. It's that easy. And it's not even hard to see why it's a pairwise equilibrium. So that's theorem one. Theorem two is going to use some of those ideas. So if you're familiar, right, with -- if you look at the greedy maximum weight matching it's a 2 approximation to the actual maximum weight matching. How many people here are aware of this? Hopefully a lot. So if you're familiar with how that works, the idea is actually going to be the same. There's going to be a bunch of details and stuff. But that's the basic idea. If you know that, this is going to seem quite familiar. So let's see why theorem two holds. So theorem two. The first thing to show, just allow us to use this greedy matching argument, is that there always exists an integral optimum solution, meaning that so by integral here I mean that every player spends all their budget on a single edge. For example, this red solution is integral optimal solution. So we can just conserve this solution instead of other optimal solutions. This is because of convexity and because of strategic complements because you're going to have more than one subject to put all of this stuff on one edge. >>: You mean social welfare? >> Elliot Anshelevich: In this talk by optimal I always mean the objective as well. In this talk I will always mean highest social welfare, yes. So highest sum of utilities. So this is the highest social optimal solution. Now let's look at pairwise equilibrium. So there's no similar result for pairwise equilibrium. Pairwise equilibrium might look like this. They might not put all their effort in a single edge. Let's take this pairwise equilibrium, compare it to the optimal solution. We better use the fact it's a pairwise equilibrium. So, for example, let's look at this edge. Conserve deviation where both of the players take all of their effort and put it on this edge. So from the green solution, consider where we take all the effort and put it on this edge. Since the pairwise equilibrium, it means that one of them doesn't benefit. Otherwise it wouldn't be a pairwise equilibrium. So which one doesn't benefit? Let's say it's this one. Let's say it's the one on the left. Okay. So what does it mean they don't benefit? It means that their utility, their utility in the green solution, which I'm going to draw like this, must be bigger than their reward they would get if both of these people put all their effort in this edge. In other words, it's the reward of the edge in the optimal solution. So this green circle, the utility of a green person, the utility of this person in the green solution is bigger than the reward of the optimal solution. Now we can do this for every edge of opt. We're never going to have the same node being pointed to by two different areas that's because of this claim, because pointed to by an arrow it means I'm spending all my budget on that arrow and I'm only spending all my budget on one edge. Because of this claim each node is pointed to at by most one arrow which means we sum this up, what do we get on the left we get the sum of all these green circles. That's the total utility of the pairwise equilibrium. Because we sum up all utilities. On the right we get the sum of these red edges. This is the total reward of all the edges in the optimal solution which is exactly opt divided by two because the reward of an edge goes to exactly two people. That's it. It's pretty easy. I'm leaving out some details, and this, but that's pretty much it. Okay. So that's why this result holds. That's why the convex stuff holds. In the remaining time what I want to do is tell you about this right here. Because I think it's pretty equal. Basically if you have concave functions, they may not be pairwise equilibria. But once you look at minimum concave, we'll always have pairwise equilibrium. I'll show you why. Just to remind you what an minimum effort games, it means all functions are of this form. There's some function H sub E could be different for every edge of min of XY. H is a single variable function, because it takes min of XY. So if -- and concave minimum effort games means all these H sub ER concave. If we have general concave functions, like square root of XY, for example, is a general concave function, there may not be pairwise equilibrium. In fact, this very simple game there's no pairwise equilibrium. You can verify with yourselves if you want. But once you look at minimum effort concave, for example, functions like square root of min XY, there's always a pairwise equilibrium. In fact, for this game there's a very simple, just spread everything out half, half. Every person puts half, half effort. The actual theorem is the following: Pairwise equilibrium always exists if the functions, if there's a minimum effort games and functions are concave and they're continuous and they're piece wise differentiable. I don't think assuming continuity and piece wise differentability is too much to ask. So now I'm going to show you how to construct the pairwise equilibrium in this case. I'm going to do basically, well, yes, I mention if we can compute it to arbitrary position, not just prove existence, if the function is strictly concave, it's unique, price of anarchy is two if it's proved. What is it about minimum effort games that really helps us? The answer it's obvious. It's the min. Like specifically what is it about the min? And the minimum effort game where the functions are something of min, and a pairwise equilibrium people are never going to have unequal contributions, right? Because suppose you're contributing more than somebody else, but your actual reward depends only on the min, you might as well contribute the same amount as them. And that's the obvious observation but it's going to turn out to be enough. >>: Question. >> Elliot Anshelevich: Yes. >>: Differentiability, it's the assumption that certainly it's not needed. >> Elliot Anshelevich: It's needed in my proof. >>: But if you have solutions, stability question, if you have solved it for some function H and HI and they converge to another one, H, isn't that implied that they converge? >> Elliot Anshelevich: Me, if they're increasing, converging to a particular function? I don't know. >>: Concave function has only -- has only so many points of ->> Elliot Anshelevich: I used it, but I haven't, frankly, truly tried to get rid of that assumption because I figured it was good enough, but I can try to think of ->>: Additional asking [indiscernible]. >> Elliot Anshelevich: It may not be too hard to get rid of. It might be worth thinking about, but not at this second. Okay. So I'm just going to give you basically -- I'm going to show you how the algorithm compute a pairwise equilibrium works on this and that should also prove to you why pairwise equilibrium also -- okay. So what do I mean by this? Now that I mentioned that the contributions of both sides are always going to be the same, because it's a min, really what I mean by here this is not three square root X, it's three square root of XY. But I'm going to write it three square root X for compactness. But it's min three square root of min XY. Here's the graph, budget. These are the budgets and the reward functions. What do we do? First thing we do is for every node V, let's compute the best, its best response. So best strategy could do. If it were able to control all the other nodes. So this is the best you could possibly do ever. There's no way you could get better utility than this. For example, for this node, this is the best thing you could possibly do. Is it would put -- budget is one. This adds up to one. But 414 from this edge and make this node match its offer. And then that's, then it will get improved square root of 414. This is the best it could do. Ever How do we figure out what this is? We can just, do it by solving convex program. This is where we're maximizing the sum of concave functions due to a linear budget constraint. Divide using convex program. But there's sort of a better more intuitive way to think about what these numbers are. And that is that well on all these edges, right, the derivative should be equal now. Of course. Because if one derivative one of the edges is higher than the other, you would just move some effort from the lower derivative edge to the high derivative edge and improve your utility. So they must be all the same. One way to compute it sort of intuitively, you look at the concave functions. You start putting effort on the ones with highest derivative, keep doing it until you come up there being the same every way. There's a few annoyances. For example, here we're assuming this person will match the contribution. We can only do that if they have enough budget. If you don't have enough budget, then they can't match the contribution even if we make them. Just doesn't have enough budget. Those are details. Okay. So for every -- so for this person there's derivative value associated with them. The derivative that it hasn't always three edges. The derivative is for every person. We computed the best edges for every node on the graph and each node has a derivative value now. You can believe me these are the best. What do we do next? We pick the node with the highest derivative value. In this case it's this one. And we fixed its strategy. This is going to be its strategy. This is going to be its effort contribution in the final solution window. And that works just fine. The problem here arises what if there's several nodes with the highest derivative value and then you can't just pick an arbitrary one. You have to be extremely careful about which one you pick. So this is what I mean by there's a crucial tie-breaking rule. If several nodes with highest derivative value, you better pick the right one and there's various things in there. But if there's only one node of highest derivative value, you pick that one and you're fine. So I won't really talk about this. I'll leave this out. But in this case this is the only one of highest derivative value. So we fix its contributions. Okay. Now we do the same thing again. We compute the best strategy for every nonfixed, for every gray node, if it were able to control all the other gray nodes. So this one is fixed. It's done. But we get the best strategy for each node if we're able to control all the nonfixed nodes. For example, so here's what it becomes. These are the new strategies. So there is one nice thing about this, right? Here, this person said I want to contribute, put one effort on here and you're supposed to put one also. And that's what happened. Well, if only that was so easy, here this is what happened, but it's not true if we just compute the best strategy for this person, if it were able to control all this, it wouldn't necessarily put one here. It might do something quite different. But there's this lemma, which really just comes from the fact we're doing this highest derivative and because we are doing the correct tie-breaking, which is it's true, not every best response of this person was the control of all of these people actually matches the contributions of the fixed nodes, but there always exists at least one. I don't think it's surprising. This is, especially once you think about the highest derivative stuff. And so we picked that one. So we can always pick contributions of these people which will match the contributions that have already been fixed. So that's what we do. And now we continue, fix the strategy of node of highest derivative. In this case it's this one. So fix the strategy and we go on. So now we do this again. Best strategy of this person if we control all the gray people, do this for everyone, we get three strategies and we keep fixing them one by one and eventually we get this solution. And the lemma is not very easy to prove. The lemma is sort of where you have to use a lot of stuff. But once you get this, I claim this pairwise equilibrium and this is actually very easy to see. Here's why it's a pairwise equilibrium. So suppose it weren't. Suppose maybe this pair of nodes would rather deviate. So it could deviate together and get better utility, both of them. Then I claim the node that was fixed first would not want to deviate. In this case this is the node that was fixed first. Why wouldn't it want to deviate? It's very simple why would it want to deviate together with this other person, why would you deviate as a pair in general? Because by yourself you can't improve your utility but if you both do something different, then my utility will actually go up. But because of this lemma and because we formed this solution, this node is actually getting the best utility it could possibly get if it could control these four people. So it's getting the best utility if we can control all of these people. Which is certainly going to be better than the best utility it could get if it controls these two people so it's already doing really well. It's never going to want to deviate. That's why it's a pairwise equilibrium. Okay. And that's it. That's the gist of why pairwise equilibrium always exists, and as I said we can then use the results for general concave function to show what the price of anarchy is at most two. So there's always a nice pairwise equilibrium. In fact, all pairwise equilibria are within a factor of two. So that's what I'm going to talk about here, and I guess I promised at some point that I would talk about this question mark. So let me talk about this question mark and then I'll mention some future work and stuff. As usual. So here's what this question mark means. I already told Nikhil about this. So this question mark means the following. In minimum effort convex games, even when a pairwise equilibrium exists, like when these three stars hold, for example, if you do best response, there may not be any path from starting point by doing best responses to pairwise equilibrium. Let me define what I'm saying. If you look at the state graph, so each node in the graph is a state, and a transition in the graph is a pair deviating, and both of them improving their utility, so the path in this is just the path in best response dynamics. So for here there are examples where there's a nice pairwise equilibrium sitting right here but you start right here and there's no path to get from here to here by using this response dynamics. It will just cycle. But not only will it cycle there's no path to get from here to here so best response dynamics are not going to converge on the fact you can't make them converge if you tell them how to deviate. But here's why I think it's an interesting question. Why it's a question mark and not just a know. If instead of doing best response dynamics, if instead of doing deviations where both people improve strictly you allow deviations where one person actually gets a better utility but the other person's utility stays the same, then at least so far I have no examples where this doesn't converge. I don't have a proof for anything. But so far every example I come up with, it always seems to converge. Somehow it says for this kind of game, being sort of a little bit altruistic, so saying I'm willing to switch as long as I don't lose anything but also if I don't gain anything. Has the same difference causes convergence to happen. Which is interesting. I don't know why yet because I have no proof. So that's what this question mark is. >>: [inaudible] this response on the [inaudible]. >> Elliot Anshelevich: No, it's a super set of the best response. More transitions. So before we only had transitions where both people had to strictly improve there's no path from here to here. Now I'm also adding transitions where only one of the people can improve and the other staying the same. So I'm adding more transition toss the graph right? There's more deviations possible now. So there's more possibility for convergence. But you're right, there's still paths which will cycle. But that's only if people choose those. Seems like there's always a path to go from here to pairwise equilibrium. So to say it one more time, it's not true that every path will now lead to a pairwise equilibrium. There are cycles in the dynamics. But it is true that there seems to always be a path to pairwise equilibrium when there wasn't before. >>: A way to identify it? >> Elliot Anshelevich: If I did I would have a proof. I don't know. So it's quite possible doing some sort of random best response dynamics where you don't take -- in the cycle you don't do the same one again, maybe that will converge, I don't know. >>: Missing something here. A cycle ->> Elliot Anshelevich: So ->>: Utilities are going up. >> Elliot Anshelevich: Utilities are going up for this pair but not for other people. >>: Right. >> Elliot Anshelevich: If it deviate ->>: Okay. >> Elliot Anshelevich: Yeah. All right. So then let me just mention, last slide, let me mention some extensions that I think are interesting open problems. So, first of all, certainly I talked about convex and concave, I think it will be interesting to look at specific class of reward functions especially ones motivated by some specific application and see what behavior is like then. Also I talked about these best response dynamics, but there are many kinds of dynamics which I haven't really looked at. Imitation dynamics and all kinds of random dynamics and no regret dynamics and all kinds of stuff. So one thing that I have looked at is general contribution games. What I mean by this is okay so far this was a graph. Each of these relationships is a pairwise thing. What if instead of a graph it was a hypergraph? So, for example, what if you have projects which with many participants, K participants get together and they all put in some effort and the reward function takes in K inputs and spits out how good the project is. So in other words what if it's exactly the same game but with a hypergraph. In that case a lot of the same things actually still work out. So all -- forget about maximum effort. Maximum effort doesn't work out. Maximum effort is kind of stupid anyway. So all the existence results work out in the same way. The price of anarchy changes unsurprisingly. So here and here it becomes K, where K is the sizable largest project. Not surprisingly. Here and here, this actually came from here, instead of two you don't even get K. You get something like, think you might be able to get K squared but you get something really annoying unless you assume strategic substitutes which is the concave, which sort of means concave or supposed to mean concave and you get two again. So you don't get K you actually get two but you have to make some extra assumptions. So that's general contribution games. And here's a few things which they seem maybe seem like minor differences but they actually cause pretty big differences to the results. One thing is I want to look at the capacities in the maximum contribution to an edge. You shouldn't be able to put all your budget in an edge. There's some maximum amount of effort you can contribute to an edge. That changes things. But you can still prove versions of results I mentioned. The really interesting thing, which is the main thing I'm actually currently working on in this area is cost functions for contributing functions. This is what I mean by this. It's what you asked earlier. And that is instead of a budget constraint, sort of a hard budget constraint, what if you have some cost function for how much effort you put in? So the more effort you put in in total the more cost you experience. So now there's a benefit for you for relaxing. Because now you don't have to pay as much cost. And that changes things quite a bit. Sometimes. I mean in minimum effort games, doesn't change things too much. If you assume everybody has the same cost function then you could still prove some result, but in general games, totally different. >>: Think of it as a loop. >> Elliot Anshelevich: That's what you do. Technically if you think about it as a loop it's slightly different. But ->>: Whatever constraints you imposed -- >> Elliot Anshelevich: Has to be here, too. But, yeah, that's what it means. It means allowing loop edges. And minimum effort it's not that bad because you're both going to be -- again you're both going to be putting the same amount and so it's going to be just a function of this and basically the amount you're putting on the edge is going to be the same for both of you, so if you have the same function then your difference because of its edge is going to be the same you can use that symmetry but if you have different functions on every edge, like different cost function on every edge or if you have over here, that's what I'm working on. So that's pretty much open. And I think it's very interesting. Finally, sort of the big thing that really changes everything is what if instead of both people getting the same reward, what if somehow you divide the reward or maybe the reward is divided proportionately to the effort you put in or things like that, that we're only starting to work on and then there's lots of different cost sharing schemes you can think of and I think it's a pretty cool question. Anyway, if you have any more questions, please ask. I'll stop here. [applause]