|
Editorial Review Book Description Updated for the first time, the classic book on why cooperation is not only natural but also the best survival strategy The Evolution of Cooperation addresses a simple yet age-old question: If living things evolve through competition, how can cooperation ever emerge? Despite the abundant evidence of cooperation all around us, there existed no purely naturalistic answer to this question until 1979, when Robert Axelrod famously ran a computer tournament featuring a standard game-theory exercise called The Prisoner's Dilemma. To everyone's surprise, the program that won the tournament, named Tit for Tat, was not only the simplest but the most "cooperative" entrant. This unexpected victory proved that cooperation--one might even say altruism--is mathematically possible and therefore needs no hidden hand or divine agent to create and sustain it. A great roadblock to the understanding of all sorts of behavior was at last removed. The updated edition includes an extensive new chapter on cooperation in cancer cells and among terrorist organizations. "This book, if read, grasped and applied, could have a profound effect." (Wall Street Journal) "A fascinating, provocative, and important book." (Douglas R. Hofstadter, author of Godel, Escher, Bach) ... Read more Customer Reviews (30)
A Beautiful Book
An idea that is elegant, simple, and powerful is a beautiful idea. That is why I think this is a beautiful book. It is elegant in its ideas and the way these ideas are presented. It is easy to read and understand with very simple math. Yet, the insights and conclusions of the book are very powerful and very interesting.
The book studies how agents behave and interact in social systems. It studies what strategies those agents might adopt and what strategies are most beneficial to the agent and to the whole social system. The results are very interesting. Another major question the book tackles is whether or not the strategy of cooperation can evolve in a social system where the majority of agents don't cooperate. The results are very interesting.
I strongly recommend this book.
Extended edition IS NOT EXTENDED!!!!
"The updated edition includes an extensive new chapter on cooperation in cancer cells and among terrorist organizations."
WHAT??I bought this, with the new preface and cover, and it DID NOT CONTAIN NEW MATERIAL IN THE BOOK.I wrote the published, and THEY DID NOT REPLY.Crappers.
Reconciling Individual Interest with Collective Interest
The goal of a model is to explain complex reality with parsimony. This means that a model is a simplification of reality or approximation to some aspect of our world. Likewise, the goal of game-model is also to explain reality by abstracting the important features of reality for a particular problem.
Prisoners' Dilemma (PD) game has represented the reconstructed reality of international politics without overriding or overarching government authority, because it does not only explain persuasively why states have suffered from the problem of non-cooperation, but also show what states should do in international anarchy.
In PD game, the more self-interest each player pursues, the more collective interest both players lose. Nevertheless, they can not stop pursuing their self-interests. Otherwise, they will be faced to the worst case. As a result, all of players do not escape from social inefficiency and eternal conflict.
Hobbes, who described human existence in the state-of-nature as Bellum omnium contra omnes, suggested Leviathanas the solution to the dilemma. However, in that states exist in "international" anarchy, the argument is meaningless. If Hobbesian state-of-nature is the reality among nations, the world is in a constant state of war.
Axelrod finds the new puzzle in here: In situations where each individual has an incentive to be selfish, how can cooperation ever develop? His goal is to explain the cooperation under anarchy. To do so, he designs a variation of PD game with R > (T+S) / 2 value and introduces the concept of time. His explanation depends on the iterated PD game (IPD)where players do not know which is the final move.
Achieving socially efficient or Pareto-superiormutual cooperation in PD game is possible if the reciprocity works as times go by and the players consider the future consequences of their present actions with foresight (the reciprocity is emphasized much more than foresight). Especially, Tit-for-Tat (TFT) strategy and the Shadow of Future play a core role in explaining cooperation.
First of all, he sets five preconditions for IPD: (1) players cannot make any threat or prior commitment (2) there is no knowledge of the other players¡¦ upcoming move, as each game is simultaneous (3) the interaction among players can not be avoided (4) there is no way to change the other player's payoffs (5) the only communication allowed is through the player's own prior behavior.
Also, the concept of a discount parameter, w, which represents the degree to which the payoffs of each move are discounted relative to the previous move, is introduced. Thus, w is the weight of the next move in the future. As times go by, value tends to decrease in negative squares.
This can be represented as the sum of infinite series.The higher the value of w, the more likely the players will meet in the future. Contrarily, since 0
From this setting, Proposition (1) is derived: If the discount parameter, w is sufficiently high, there is no best strategy independent of the strategy used by the other player.
Also, Axelrod runs two computer tournaments by inviting the top game theorists, and reports that Rapoport¡¦s TFT was the best strategy. Although 15 programs in 1st tournament & 63 programs in 2nd tournament were submitted, TFT won all of them. TFT got the average score of 504.5 in 1st round & the score of 434.73 in 2nd round. Why?
He divides the reasons into 4: (1) TFT avoids unnecessary conflict by cooperating as long as the other player cooperates; this is being nice and never being the first to defect (2) TFT has no hesitation in retaliating in the face of the other's defection; that is, TFT does not wait to see if the defection was a mistake, as defection is to be punished immediately.
(3) TFT is open to forgiveness after responding to a provocation; in this way, a TFT player do not wait to reward cooperation (4) TFT has the clarity of behavior, so that the other player can adapt to TFT pattern more easily than others (TFT program length was the shortest). Thus, TFT could do well over a wide range of environments, against both nice and defection strategies.
Next, the stability of cooperation based upon TFT is discussed. He explores the relationship between a native population uniform strategy & a newcomer strategy. He assumes the existence of a native population employing strategy B & a newcomer using strategy A.
If the players interact with each other one at a time, the expected utility of the newcomer with A might be higher than the expected utility of one of the native population. In that case, A is said to invade B.
Otherwise, B is said to be collectively stable. When p is the frequency of a newcomer interacting with other newcomers, the condition of invasion by newcomers is:
(p)*EU(newcomer|newcomer)+(1-p)*EU(newcomer|native)>EU(native|native)
If p is between 0 & 1, newcomers¡¦ strategies can invade the strategy of the native population and vice versa. So, can TFT invade All D? Can All D invade TFT? Assume w = 0.9.
(p)*EU(TFT|TFT)+(1-p)*EU(TFT|All D)>EU(All D| All D)
(p)*[R/(1-w)]+(1-p)*[S+(w*P)/(1-w)]>P/(1-w)
(p)*[3/(1-0.9)]+(1-p)*[0+0.9*1/(1-0.9)]>[1/(1-0.9)]
(p)*(30)+(1-p)*(9)>10
21p+9>10
21p>1
p>1/21
(p)*EU(All D|All D)+(1-p)*EU(All D|TFT)>EU(TFT|TFT)
(p)*[P/(1-w)]+(1-p)*[T+(w*P)/(1-w)]>R/(1-w)
(p)*[1/(1-0.9)]+(1-p)*[5+0.9*1/(1-0.9)]>[3/(1-0.9]
(p)*(10)+(1-p)*(14)>30
10p+14¡V14p>30
10p+14-14p>30
-4p+14>30
-4p>16
p<-4
Thus, when the shadow of future is very strong (w = 0.9), TFT can invade All D if there is more than 1 TFT invader for every 21 All D natives. It takes so few. Also, All D cannot invade TFT when w is sufficiently high (strictly speaking, the critical value of w for TFT to be collectively stable is 2/3).
From this, he derives additional 6 Propositions. The most interesting one among them is Proposition (6)The strategies which can invade All D in a cluster with the smallest value of p are those which are maximally discriminating, such as TFT. This means that cooperation is possible even in the world of All D, as long as small clusters of discriminating invaders with TFT have a small proportion of interactions in the Hobbes state-of-nature.
More interestingly, he shows that cooperation could emerge even without friendship. Let-and-Let-Live system (i.e., the static nature of trench warfare) might be considered as the prototype that small TFT interactions invaded the strategy of All D native population on war. However, Axelrod also mentions that the stability of cooperation based upon the reciprocity can be impaired by (1) the rotation of troops (2) the artillery less dependent on reciprocity for its life than infantry in trench (3) the raids.
However, pointing out that Chapter 4 is about interaction among human beings who can evaluate the reciprocity and respond to it rationally, he argues that such understanding by the participants is not really necessary for cooperation to emerge and prove stable.
Therefore, he argues that (1) cooperation is possible without morality or foresight as shown in the relationship between crocodiles and crocodile birds (2) the patterns of unconscious responsiveness of bacteria or organism might lead to the cooperation based upon the reciprocity (3) the evolutionary process depends upon individual advantage (not benefits to whole group), which unintentionally leads to the cooperation based upon the reciprocity.
In addition, he suggests four advices on how to choose effectively under a given strategic setting: (1) Do not be envious; TFT never wins head to head, so players must realize that an IPD is not a zero-sum game (2) Do not be the first to defect so long as the future remains important, based on Proposition 1 (3) Reciprocate cooperation and defection but begin with cooperation (4) Do not be too clever; be clear about your strategy so others can figure out what you are doing. Again, TFT is the strategy which satisfies all of the advices.
Especially, five ways on how to promote cooperation is discussed in aspect of changing the strategic setting: (1) Enlarge the Shadow of the Future by making the interactions more durable and more frequent (2) Change the payoffs; The change of payoffs determines the incentives of behavior (3) Teach people to care about each other (4) Teach reciprocity; Do not forget the negative effect of All C that might spoil the other players (5) Improve recognition abilities; Accumulate the credibility of reciprocity through good history of interactions.
Finally, the social structure of cooperation is discussed. The social structure influences on how the evolution of cooperation can begin. The influence might constrain or facilitate cooperation, or make the evolutionary process of cooperation dynamic. Namely, the relationship between the social structure and the cooperation in IPD can be understood as being equal to the relationship between the culture and the institution.
As the institutional performance depends largely upon culture, so the speed and the range of the evolution of cooperation is determined greatly by the social structure such as labels, reputation, regulation, and territoriality. Fist, labels (i.e., stereotype) might decrease the importance of the benefits due to mutual cooperation. This is related closely to the debates on the distributional effect of collective interest.
Second, the importance of reputation as a bully might delay the speed of cooperation. Third, relating to regulation, the government in here is not Leviathan, but a player interacting with the governed on compliance and flexibility. In this case, the efficiency of the exchange of flexibility with compliance determines the evolutionary process of cooperation among them.
Finally, in that the territorial system (i.e., positional picture) influences the way the players interact with each other which determines the course of the evolutionary process, territoriality as the social structure matters.
Axelrod concludes: (1) Cooperation has staying power but the biggest problem is getting cooperation started (2) Ratchet effect: Cooperation is successful incrementally, as clusters of cooperation build upon clusters cooperation (3) Cooperation is a rational possibility, even without a central authority, as long as the future is sufficiently important (4) Reconciling individual interest with collective interest is possible by TFT.
Recently, Fearon (1995), Morrow (1999) and Powell (2005) argue that one of the reasons why the Pareto-inferior outcomes such as ex post costly wars have recurred is due to the commitment problem. Morrow (1999: 92) maintains, Commitment is a problem when actors' incentives change over time.
Although Axelrod and they consider the concept of time seriously, their conclusions are totally different. While they focus on explaining the cause of war in PD (i.e., In PD, one player's commitment to C can not be believed by other player), his interest is placed on explaining the cause of cooperation in PD (i.e., what matters in PD, is not unilateral commitment, but mutual learning effect by TFT).
However, Axelrod seems to be superior to them, because he might explain both war and cooperation with the level of w. But, Fearon, Morrow, and Powell might suffer from selection bias, because the commitment problem explains only wars.
References
Axelrod, Robert (1984). The Evolution of Cooperation. Basic Books.
Fearon, James D (1995). Rationalist Explanations for War. International Organization 49 (3): 379-414.
Morrow, James D (1999). The Strategic Setting of Choices: Signaling, Commitment, and Negotiation in International Politics. In Strategic Choice and International Relations, ed. David A. Lake and Robert Powell. Princeton, NJ: Princeton University Press, pp. 77-114.
Poundstone, William (1992). Prisoners' Dilemma. Anchor Book.
Powell, Robert (2005). War as a Commitment Problem. International Organization (forthcoming)
The evolution is just beginning
Amidst the glowing 5-star reviews I feel the need to interject some concerns and outright criticisms of this book, although I do recommend its reading.Certainly the book provides a relatively good starting point in a very complex area, but it should not be construed as the final word on a much more complex subject.I do think that R. Axelrod provides an excellent, if at times overbearing, presentation of how game theory, specifically in the realm of an Iterated Prisoner's Dilemma (IPD) scenrario, can explain a number of historical as well as daily situations.He goes on to expound on some good generalizations on how people might act to inspire more cooperation and these generalizations, if implemented by people (groups, governments, etc.), might result in a more cooperative world. For this, there are good things to say.But I would caution not to create from this any utopian potential for the real world.
The `winning strategy' of TIT FOR TAT (TFT) works because it starts out `nice' but it retaliates immediately if someone else does not `play nice', too.So this is not vision of world cooperation.It is a realistic vision of maximizing cooperation under specific conditions which he covers relatively well towards the later chapters.Ultimately, the natural take away is a hopeful view of a potentially more cooperative environment, with perhaps a bit of forgetting that unkind retaliation is an integral part of his winning `cooperative' strategy.
Here are some aspects which Mr. Axelrod alludes to but somewhat minimizes, in my opinion, in their impact on TFT's potential for success in real world interactions:
a) The IPD strategies, which were submitted by experts from around the world, were submitted to computerized testing to determine which strategies `win' the most.This is based on an established point system that awards different points for different actions by two players.(This is summarized in one review already, so I will not repeat it here.)While this makes the playing of the game easy and consistent, it does not reflect the real world conditions which often exist.For example, if, in one turn, one player `defects' and the other `cooperates', the point system says the defector `wins' 5 points and the cooperator gets nothing.Are all defections `equal' in the real world?Or are some interactions far more important than others, so therefore cooperating on small things but defecting on bigger things might result in different outcomes than would ever be accounted for in the point systems used in these studies?Intuitively, a major defection may have far more significance (i.e. point value) than a minor one, but to accomplish his analysis, Axelrod's point structure is always the same.
b) Related to the above, are all joint cooperative efforts or joint defections `equal' as implied by the point values?(When two strategies cooperate, each is awarded 3 points.When each defect, they each get 1.)At the time this book was being published, an interesting `real world' example was playing out which would question this very assumption.At that time, Reagan was building up the military and the Soviet Union was `matching' the build up.You can either view this as mutual defection or mutual cooperation, depending upon whether you view cooperation as always a `positive' thing.In any event, the results of these mutual actions were NOT equal.The US had far more resources to invest in the game so each time `points' were awarded, the US actually gained more than the Soviets who eventually had to stop playing. Consistent point values simply do not account for this, unless one wants to interject additional elements not presented in Axelrod's work.
c) Related to point b), the IPD study presents things from a vantage point that each actor starts out from an equal footing and therefore the only functional question is whether one strategy consistently wins in such a way as to keep near the top of the point standings.In reality, rarely does any person or group begin interaction on an equal footing.What I am trying to raise is not the same as his discussions in Chapters 8 and 9 on the strength and growth of various strategies in a world starting out with many different strategies.In those scenarios, some strategies die out because they stop having sources of points to take from others.But what if each strategy starts out from an unequal basis, some having significantly more `capital' to expend and coupling this to a more realistic scenario where each interaction is not free, but each has a cost to the participant (somewhat like putting your bet down before you play poker - you may win or lose, but if you lose your position is not the same as when you started, it is less than when you started).In such a scenario, which is more realistic in terms of how companies, governments and even people interact, the results would be materially different.So the starting premise of Axelrod's IPD scenario and point schemes may be so well constructed as to make for interesting descriptions of some specific and even common interactions, but it may be too well constructed to be able to be extrapolated to many more complex situations.
d) Another weakness is the assumption that all defections are of the same magnitude.If, in an otherwise nice political campaign (ever seen one of those?) a small, third party candidate launches an attack ad and the attacked major party retaliates, is the retaliation equal?Or, again, does their starting point allow them to annihilate the attacker?(Since I have never seen a positive political campaign, you can interject any other similar scenario which might actually come up in real life!)The `equal points for equal actions' premise is inherently flawed.Yet this is the basis of much of the book's conclusions.
e) What is the end result of a universe where everyone uses a TFT strategy?This is only marginally considered, at best.While the issue had been in my mind throughout much of the book, it was not until fairly late that Axelrod makes clear, albeit briefly, that in any set of two player interactions, TFT will at best result in the same total number of points as the other player and, for a number of reasons, probably slightly less.In other words, the person, company, group or government using a pure TFT strategy must be happy being close to the best in whatever the interaction.Personally, I am fine with that, thank you!But to assume that everyone is and that nobody will come up with a disruptive strategy that, perhaps only for a time, garners more points but winds up putting them on top of the heap in whatever competition may be in play.If the end game is the Superbowl, being second is not gratifying.Ask Philadelphia.
f) Finally, I found some of the descriptors attached to the strategies interesting, especially in light of Axelrod's ending recommendations on cooperation.Some strategies which attempted to use planned defection as part of their strategy were labeled as `meanies' while other strategies that started their initial moves with cooperation were labeled as `nice'.I find it interesting that a `scientist' would use such descriptors to classify things given that they bring with them connotations.Even TFT could be construed as a `meanie' since one of its logical outcomes is that it cooperates once, and once only, only to defect from then on based on the other parties action.Are all non-cooperative interactions except one inherently more `nice'?This is not important in the scope of my real concerns, but I always sense a red flag rising when an otherwise objective presentation resorts to affect-linked labels to make part of its case.
All of the above is NOT to imply that I think Axelrod's work is wrong and should be thrown out.I simply felt that, unlike the rest of the glowing reviews, Axelrod's work does not go far enough and leaves some gaping holes yet to be explored.Read this book, but keep your eyes and mind open for the flaws which seem to be thinly covered with papier-mâché.
How mutual hostility can evolve into cooperation.
Without question, the case studies in this book have applications in biology, sociology, international relations, economics and business. The basic question put forward is, "How is it possible, that in an environment of mutual hostility where acting selfishly will lead to gain against your opponent(s), cooperative behavior between the antagonists will emerge and become the dominant long-term behavior?" It turns out that it is easy to see how such behavior can emerge, even in hostile battlefield conditions. In fact, cooperative behavior has distinct evolutionary advantages.
The solution is found within game theory, in particular the situation known as the Prisoner's Dilemma. Two people, (one and two), who jointly committed a crime are arrested for the crime and placed in separate rooms where they cannot communicate. The police interrogate him or her separately and offer each individual a deal. If they defect and testify against their comrade, they will be given a reduced sentence. In this situation, there are four possible outcomes:
1) Neither defects - both go free, each is considered to have earned a positive reward.
2) One defects and two does not - one is set free and two serves a long sentence.
3) Two defects and one does not - two is set free and one serves a long sentence.
4) One and two both defect - each serve a reduced sentence.
In the problem, reward values are assigned to the results, and typical values are
1) Both one and two are both assigned a value of +3.
2) One is assigned a value of +3 and two the value of -5.
3) Two is assigned a value of +3 and one the value of -5.
4) One and two are both assigned a value of -1.
It is clear that each prisoner wants to avoid the situation where they are the only one who serves time in jail. Therefore, if this event will only occur once, then option four will be the result and cooperation will not take place.
However, if both prisoners have the potential for a future relationship, where that relationship has the real potential for rewards for cooperation and punishments for defecting, then option one can emerge. The best demonstration of this is what took place in some sectors of the western front in the First World War. When the same units faced each other for extended periods of time, a live and let live policy emerged on both sides. Each side adopted a strategy of not engaging in lethal force, unless the other side did. When required to expend artillery ammunition to demonstrate aggressiveness to superiors, they would shoot the same target at the same time of the day. Since their firing was predictable, soldiers on the other side would know to avoid that area and in fact would often climb out of their trench to observe the explosions.
There were instances where German snipers would demonstrate their prowess by continuing to hit the same position on a wall until they made a hole. Therefore, even though superiors admonished the soldiers to continue to kill the enemy and both sides had the capability, the fact that they had a lengthy relationship allowed the cooperation to occur. These phenomena did not take place in regions where units did not face each other for extended periods.
The first chapter describes tournaments, where computer programs competed against each other by defecting or cooperating and the scoring is similar to that of the Prisoner's Dilemma already mentioned. What emerged as the most successful tactic, even when the results of the first round were incorporated into the second round, is the TIT FOR TAT. This strategy is very simple, cooperate in the first round and for each successive round, do what the opponent did in the previous one.
I was fascinated by these results and it was easy to see the obvious implications for relationships of all types. For cooperation to occur, all that is necessary is that there be the expectation of a continued relationship and the potential for future rewards/penalties. What makes it especially interesting is that no appeal to morality, ethics or any other abstract concept need be made. The behavior occurs as a consequence of an increase in the long-term gain for all parties.
Published in Journal of Recreational Mathematics, reprinted with permission.
... Read more
|