Analyzing Games in Maker Protocol: Multi Agent Influence Diagram Approach by Sudipan Sinha

Transcript

uh uh greetings to everyone out here uh and today I'm going to talk about one of the most interesting topics in the market right now which is all around Economic Security and economic risks with igen lay coming into the narrative to build the crypto Economic Security layer of ethereum uh this has been one of the hottest topics but what's what's going on at chain risk so at chain risk we're trying to prevent defi biggest economic exploits whether it be the D5 which we've seen in the past and or the D5 which is right now getting built on top of lrts all the money markets which are getting built on top of lrts right so who are we so we are a passionate team of researchers who have experienced in some of the top most companies in the

world starting from deche Bank NASA uh Walmart um and some of the top colleges of the world but on top of it uh you know we've been really working hard with some of the protocols right here so we are already working with compound on their economic risk management uh but on top of that right now as I speak we are also launching our Moro walls for angle um angle is a stable coin protocol so we're launching their Moro walls and along with the collaboration on Renzo so pretty much uh uh uh what we are going to talk about right now here is recently we released a paper this paper is about the economic security of maker so anyone here will be who will be willing to read

this can just scan this QR code and just go down to the paper and scroll as I speak uh but what is this all about right so if we speak uh how we actually visualize the entire defi industry to be then it's moreoever just like a game which is happening amongst the governance the users and the protocol parameters which have been set in place right so uh I'll be talking about something called multi-agent influence diagrams which had been introduced used by two of the V uh very famous mathematical researchers called Coler and milch which is now being very deeply worked on uh by the Google Deep Mind team as well so why even multi-agent influence diagrams right so there's an

open research statement research problem statement right now on ethereum uh which says about the formal verification of the game theoretic mechanisms which are there in D5 protocols it's an open research problem which has yet not been solved the second is uh you know you really have a lot of incentive mechanisms in these D5 protocols how do you optimize those incentive mechanisms so that any agent which is interacting with the D5 protocol agent means users or any Liquidators or any arbitragers how are they playing even an equal game which means a rational and an honest game uh throughout whatever transactions they make and the third is how are we able to get or conquer a Nash

equilibrium for each and every D5 protocol which means an an optimal strategy and an optimal set of parameters for the protocol so the protocol is stable in all market conditions so what what WTF are multi-agent influence diagrams this a very heavy topic to even get a conversation into but multi-agent influence diagrams can be directly related to agents so in defi in all of the economic risk management stuff we do we something we do something called Agent based simulations so all of the interactions which are happening with a protocol can be broadly classified into three agents Liquidators arbitragers and or borrowers and lenders right so all of

these agents are interacting within a protocol and all of the different interactions can be classified as different nodes in a graph now there are three types of nodes which are possible the number one is Chance variable right so um what's the possibility or what's the probability of a particular event happening let's say if let's say today rain occurred then what's the possibility that we could get a crowd here in E Dubai then the possibility of rain let's say 5% or or even a 95% uh we've seen all of that in Dubai this time uh what's the possibility of that that's something which is called a chance variable which is sort of an external environmental thing which is

affecting you the second is all about decision variables right so how are you actually quantifying the decisions which you take so your decision to come to here in this case of a rign is a decision variable which is directly getting impacted by the probability of the chance variable and the third is utility variable what so what's what was my mindset when I was trying to come here that's to deliver the speech to you guys and introduce you to the concept of multi-agent influence diagrams so I want to deliver my speech in the best way possible right so the utility in case of defi could be to maximize your strategy maximize your yield inputs or maximize your yield outut fots right so pretty much that's what the three variables of multi-agent influence diagrams looks

like I'll share this deck with all of you guys I'll have a conversation with Patrick after this to share uh so that you guys can dive into the major mathematics part of it but let's Deep dive into the uh multigen influence diagrams with an example right so what does a multigen infuence diagrams looks like so let's say uh there's a girl called Alice who wants to build her own pasio uh in front of her garden but then uh there's this guy Mark who's just stay who just stays next door he has uh let's say a big tree in front of his house which is blocking the view of the peso with a very beautiful beach which is which is outside so what can Alice do right so Alice comes so Alice doesn't care about the world she's an absolute

jenzi she just goes ahead and says that hey I'll just poison the tree which is there and to poison the tree there'll be an effort which I need to do then the tree might get sick and finally the tree might die right but then uh you know Max thinks that if my tree is dying so max cannot see that uh the tree has been poisoned but he can see that the tree tree is dying right uh so he can directly call a tree doctor to call a tree doctor there'll be a cost which he needs to incur that will again directly in uh like affect whether the tree dies or not or whether the tree lives or not and finally all of this will affect alysa's view from the pasio right so if you see the graph over here so it is

clearly showing that poisoning the tree through the dotted line is directly affecting the pasio uh Max might call a uh a tree doctor uh he can cure the tree that will come at some cost to poison the tree and to monitor it um Alice would always need to uh sort of do an effort and check whether the tree is dying or not right so within this if you see so which one is the chance variable which one is the decision variable and which one is the utility variable right so the decision variable are the top three things squares which is telling about the poisoning of a tree calling a tree doctor or building a pasture these are decisions which you want to do uh what's the utility variable the utility variable are the things which you are incurring so what's the cost of it you

want to minimize the cost of cost you want to minimize your effort and finally uh you know you want to build the view so the view also comes in as one of the utility variables you want to maximize The View right and finally uh what's the sort of chance variables in this the biggest chance variable is of course the tree getting sick and finally the tree getting died right so what's the probability of that so that that's basically shown in ovals so how are we building multi-agent influence diagrams for maker Dow right so if you see the makeer protocol then uh there are a lot of governance uh and user activities going on so let's say we U you know quantify all the governance uh Dow as a set of Agents you know set set of Agents from G1 to let's say GN and similarly

the users of the protocol let's say up P1 to PN and then how are we building up the model right so let's look at it very closely so it's a uh it's a very interesting thing to look into number one is what are the chance variables market conditions collateral Dynamics demand for die and the security risk security risks are inherently coming down from the smart contracts all of the rest of the three are basically Market uh conditions which are coming here uh all of these are finally affecting your decision variables decision variables over here would be risk parameters which the governance and the decides uh which pretty much means uh it directly affects the amount of debt the user can borrow as well which is also coming down as one of the decision variables and using all

of this we're trying to get two utilities and maximize two utilities number one is the profitability and the second is the sustainability of the protocol all of this will come at a risk and reward uh ratio so you have to so this is basically a trade-off right so whenever you want to increase the top lines there'll be an increase in the risk and then you need to also uh you know sort of manage the risk at the same time so that you don't get hacked and all of that so this is basically how the ma looks like in case of maker Dow uh you guys can go through a research paper to uh look into the depths of the uh modules and details uh but that's pretty much it uh how what the problem does made solve for maker out right so we're coming up with a uh next paper in eth

Belgrade which uh clearly gives you the outputs of it so what are we doing over here number one is calculating the utility right so you there were two utilities one is the profitability and the second one sustainability how do you maximize both of these using strategies so how can we find the optimal strategy Sigma so that at any given point of time my protocol is maximum sustainable and at the same time maximum profitable and secondly is to evaluate the Nash equilibrium now what do we mean by Nash equilibrium uh you know it it says that a strategy profile is a Nash equilibrium if no agent has the incentive to deviate from the strategy specified from the profile by him as long as the other agents do not Dev from the specified

strategies right which means that all of the players in the ecosystem all of the agents are playing evenly they do not have any added incentives and all of them are not only rational but they're all honest which means that nobody has a gain over somebody else and if they have it they have it already rationally which is the game of the entire D5 space that somebody makes money somebody loses it right so that's pretty much what we are doing for maker Dow um happy to take up some questions on this don't let your risk program take a said back you can quickly scan our Twitter you can quickly go to my Twitter as well I usually post a lot of uh research content on this so yeah happy to take up any questions if there

yeah yeah so um uh there are basically three forms of Agents here one is of course the uh uh Liquidator the second is the arrer and the third is the borrower so we are modeling them especially using all of the main net data which is uh present below and we're using reinforcement learning models to all to model all of these and then finally run all of these simulations in our platform uh where we have built a huge scalable platform on top of revm where we are running multicar simulations at scale so it's to give you the con confidence intervals and converge that montic Carlo simulations to the different set of parameter values so that you're able to finally get into the Nash equilibrium fine cool any other questions or I think we are already on time we'll wrap it

up cool then uh pretty much excited uh to have a conversation with with you guys outside