Ensemble de rappel
Ensemble de rappel
Ensemble de rappel
Ensemble de rappel
Ensemble de rappel
Ensemble de rappel
Ensemble de rappel
Lecture – 19Machine Intelligence
(Refer Slide Time: 00:15)
Welcome back for the next session. So were discussing about the economic origins ofdigital supply network and we talked about different laws, which comes along withdifferent economies. I mentioned about the Huang’s Law, but I have not discussed it.So, I request you, if you have time just have a look at it. I will continue with mydiscussion on the technology of the future.(Refer Slide Time: 00:54)
I just briefly talk about machine intelligence. So, just put a caveat up front this this ismy version of what is machine intelligence and this suddenly cannot answer, maybewe are not going to discuss the algorithms, but we will try to get some idea aboutwhat does this machine intelligence all about. So, in fact, you can actually seedifferent flavors of it.
So, it goes back to the work of Alan Turing, but the modern history of machineintelligence goes back to 1956 and if you recall, when we were talking about thecomplex supply chains, we mentioned that whether machine intelligence can help toreduce the complexity of those supply chains. So, this brings us to the setting wherewe have agents and we are looking for intelligent agents, and then there is anenvironment in which they have to decide something.
So, when we talk about agents, we as I mentioned that we assume that they areintelligent and the environment is a bit dynamic. We can call it variable and uncertain.Now given this dynamic and uncertain environment, agents would have someobjectives to achieve.
You can think of a pretty simple example that the demand is fluctuating and I have todecide how much inventory to keep and my objective may be to maximize the profitor minimize the cost. So, humans are intelligent to the extent that our actions can beexpected to achieve our objectives. Now when we extend this argument, so can Ireplace this agent with the machine.
So, instead of calling it intelligent agent, let me think whether I can actually callintelligent machine. Can the same thing be extended? So, machines are intelligent tothe extent that their actions can be expected to achieve their objectives. Nowmachines if they achieve their objective, because I think this goes to very deeperinsights like whether machines are conscious, whether machines actually can havethat hedonic feeling.
So that normally we associate with humans. Machines are intelligent to the extent thattheir actions can be expected to achieve their objectives. From the perspective ofmachine intelligence, this objective is not correct. What is correct is, in fact, I have
taken it from a recent book by Stuart Russell, which is Human Compatible. So,machines are beneficial to the extent that their actions can be expected to achieve ourobjectives.
So that is in fact defines what actually should be machine intelligence. So, even if youdelegate some actions to the machines, they should be beneficial to the extent thatthey can achieve our objectives, but is it enough. Because human beings, even when Igive you a typical inventory optimization example. Even if you recall the EOQ kindof simple formula, it has been observed in practice that the managers actually do notoptimize.
So, they actually satisfy. They actually would have lot of biases when they actuallymake the decision. In fact, the whole area of so if you have heard about this book byan economist, Nobel laureate Daniel Kahneman,, Thinking Fast and Slow. This booktalks about lot of biases and heuristics, which we use when we decide. So, whetherthis objective is good enough or whether we should do something else.
This brings us to the modern paradigm of machine intelligence. The machines arebeneficial to the extent that their actions can be expected to achieve our optimalobjectives. It is not just the objectives; it should achieve our optimal objectives. Thisgives the paradigm in which we want to look at machine intelligence. So this, so fortime being, let me make it clear that this only gives maybe the epsilon part of how themachine intelligence should look like.
We are not going to discuss the algorithms. I will give you some used cases wherepeople have used the machine intelligence to as part of the digital supply network. ButI am just giving you the flavor of what actually should be machine intelligence.Machine learning could be a subset of machine intelligence, which may be datadriven. But the whole objective may be much larger.(Refer Slide Time: 06:34)
I will give you one setting which coming from Game Theory, and this would beinteresting to see whether how to look at an agent and how to look at environments.So, let me start building this example and this will give you lot of insights about it andwhat actually happens. I will take some time for this. So, if you see this example, youcan see that there are two players given; player 1 and player 2.
Now player 1 is descending in an environment in which the outcome is alsocontrolled not just by player 1, it is actually controlled by player 2 also. So, you canactually see player 1 controls T or B. We can call it top or bottom and player 2controls left or right. But the payoff, which is the first payoff is for player 1 andsecond payoff is player 2.
So, if assume player 1 chooses T and player 2 chooses the outcome is (10, 10). Player1 gets 10 and player 2 gets 10 and same thing will happen if T, R. So, it becomes (0,11). So you can actually see the whole payoff matrix is actually a function of jointaction of both the players. Now we make lot of assumptions around it that both theplayers are assumed to be rational. They are strategic and self-interested.
When I say rational it means the players are intelligent enough to know what actuallyis the best action for them and they are intelligent to such an extent that they actuallymake out, so assume that this payoff matrix is known to both the players. But theycannot control the action of the other players, because both the players are assumed tobe self-interested.
So, they are looking for larger payoff for themselves. They are also assumed to be aswe mentioned that rational. It means that player 1 can infer what is the best choice forplayer 2 and player 2 can do the same thing. So, they are not intelligent enough to justknow what is the best action for them, they are intelligent enough to infer what couldbe the best action for the other player and they are strategic.
It means that they want to move the outcome towards that in which they get a largerpayoff. So now I hope you are getting. Player 1 and player 2 could be the agents inthis case. The payoff matrix is the environment. But even if they have informationabout the environment, they cannot determine what is the best action for them.
They cannot control the outcome because the outcome is not just in control of player1, it is also in control of player 2 and vice versa and let us make that assumption thatboth the players make the choice independent of each other and withoutcommunication. Player 1 is not communicating to player 2 that what I am choosingand vice versa. So, player 2 does not communicate to player 1 what player 2 ischoosing.
Now given this setting what is the best choice for both the players. Now let us seehow this whole thing actually works out. So, player 1 is not communicating to player2 and player 2 is not communicating to player 1. So, what is what normally you infer?Will they actually converge on T, L or B, R? So, let us see that.
This gives you the idea about what I am saying in the previous one that how themachines should think about or you can say the machine should be rational to lookwhat is the best objective for us. So, let us start building the case. Player 1 does notknow what is the best choice of player 2. So, let us assume that player 1 assumes thatplayer 2 is choosing L.
So, in that case I would respond by playing B and when he chooses R again, hechooses B and same argument you can actually see this guy will choose. If theycannot coordinate their actions and there is uncertainty about the choice of the otherplayer the equilibrium or the optimal output is B, R which gives them a payoff of (3,
3). When there is a better outcome available, which is T, L which gives them a payoff
of 10, 10. What does that mean? It means that even if there is a better payoff the self-interest leads to an outcome which is actually worse for both of them. Now the
question here is if you want to machine make the choice, will you still prefer B, R orwill you go to T, L and there is uncertainty about the choice of the players.
Now let me extend this argument if so I think when I talk about so this whole gametheoretic part is in fact also called as multi agent setting and we are actually sayingthis because as we talked about the digital supply networks, you would actually seeintegration of so multiple firms multiple consumer and all these consumer and agentsare assumed to be rational.
They are assumed to be self-interested and they are assumed to be strategic. what doesthat mean? It means that, you would actually need the modeling tool of game theoryor multi agent settings to understand the dynamics which happens as part of thesupply network. So in fact, this B, R is actually called as Nash equilibrium.
Nash equilibrium is in fact, John Nash paper was 1950 and very famous movie on thelife of John Nash, A Beautiful Mind. So, if you have time just watch that moviesometime. What it actually is doing is we are converging on an outcome which isinferior to both the players. This is something which, so which gives you what isoptimal.
So, in this case, if they could coordinate, there is an outcome which is T, L. If they arenot coordinating, they are converging to B, R and this whole idea of self-interest orstrategic or rational and Nash equilibrium, this is pervasive. I am pretty sure if youstart thinking you can actually realize that in our day to day life, this is somethingwhich we observe most of the time.
Only thing which we can achieve, there is an outcome which is 10, 10. So whether themachines actually allow us to achieve our objectives or will they actually allow us toachieve our optimal objectives. So, the first the third line, if you see may converge onNash. The fourth line will allow us to go to 10, 10. Now I extend this argumentbecause the same argument we need to use in blockchains also.
Now let us assume that, if they communicate with each other that assume player 1communicates to player 2 that I am going to play T and player 2 also communicates toplayer 1 that I am going to play T. Now when the game is played, so this is just acommunication. They are not actually going to penalize if you deviate. What willhappen? Both the players when the game is played, they try to cheat the other player.
This has been observed in lot of class experiments also. I have played lot of theseexperiments. So, player 1 assuming that the other player is going to play L, if I play BI get 11. He tries to go from T to B. But the same logic should be used by this player.So, both will actually converge to the same B, R even if there is a communication andthat communication in game theory is called as cheap talk.
You would realize in the next discussion on blockchains, the main idea of blockchainis to actually avoid this cheap talk so the communication becomes credible andtransparent. So, the logic is mainly game theoretic and the idea is to avoid this cheaptalk. With this l will go to the blockchain part, which I will do in the next session.Thank you.
Nous enverrons les instructions pour reinitialiser votre mot de passe à votre adresse e-mail associée. Veuillez marquer votre adresse e-mail actuelle.