Some have argued that without religion there would be no adherence to any morality. Could this be true? Does morality necessarily owe its beginning and continued existence to a belief in the supernatural? The position of this paper is that the answer is 'no'. The argument type will be rather different from the standard ones used in philosophy in that I will attempt to borrow the methodology of Complexity Theory.

It is true that I am taking a position that should not require much of a proof, since it is only the position that ethics could have developed without institutions. But there are many who would not buy the standard philosophical arguments. So I would like to present one more argument that religion and institutions are not necessary for the maintenance of ethics and morality. This argument will attempt to address the concern directly, by proposing alternative forces for the creation and maintenance of morality and showing that these forces would be sufficient for that result.

I will not be arguing that ethics and morality were in fact maintained and developed by these alternative forces, but instead that they could have easily emerged without institutions. However, in addition, I will also show that if it were not for institutions it would be virtually inevitable (no pun intended) that ethics and morality emerge from natural mechanisms and become entrenched within our societies.

One might think that this a thesis more fitting in a sociological context and not a philosophical one, but if 'the nature of morality' is a topic for philosophy, then 'how morality and ethics could have developed' might be important to the discovery of that nature. A concept's role, origin and function may give clues to its meaning.

First, a little about Complexity theory and how it has been used to address philosophical problems in the past. In particular complexity theory has something to say about the prisoner's dilemma.

The classic prisoner's dilemma problem starts like this.
After committing a crime, you and your buddy are caught by the cops and separated to be grilled. You are both offered the following deal.
If you squeal and your buddy doesn't, you will be let off scott free.
If you don't squeal and your buddy does, you will get 20 years in the pen.
If neither of you squeal, the detective still has some minor violations that can be pressed or trumped up, and you will both get 6 months.
If both of your squeal, then you will both get 15 years.

What is the rational thing to do (from the point of view of self interest)?
It would be nice if you and your buddy could talk it over and act in unison then both of you would get 6 months by agreement to both clam up, but you are not allowed in the same room and there is no legal contract running in the background.
So consider the cases.
Suppose you know your buddy is going to squeal:
Then it can be argued that it is best for you to squeal for then you will get 15 years instead of 20
Suppose you know your buddy is not going to squeal:
Then it can be argued that it is best for you to squeal, for then you will be set free instead of getting 6 months.
So, in either case it is better for you to squeal.

So two people doing what is rational and in their self interest end up with a poor outcome, even though had the choice been put to them collectively they would rationally have decide not to squeal. In fact, if they knew beforehand that this situation was going to come up and if it were possible it would have been rational to lock each other into an iron clad contract that would have made squealing impossible for both.

But you might suggest that if they only trusted each other, things would be different. However, the reasoning is not dependent upon non-trust, for even if you had trusted that the other will not squeal and you know he is totally trustworthy, it is still to your advantage to squeal.

Of course, if you care for your buddy as much as you do for yourself then a different outcome would result. However, for any caring less than 100%, there is a way to set up a dilemma problem.

As abstract and as limited in context as the problem seems, it can be convincingly argued that many of the problems facing individuals and civilization's continuing existence are traceable to problems analogous to the prisoner's dilemma. Anytime cooperation is needed but there seems to be no mechanism that will enable it, we have the prisoner's dilemma as a prime suspect.

However, most of the realistic prisoner's dilemma kinds of problems have an extra feature. We are often in the same kind of situation with others, repeatedly. So consideration was given to the following question: Suppose we know that we will be in the same situation over and over again with the same person, would the 'rational thing' to do be different given this context of only self interest?

A contest was devised where the contestants were to play a prisoner's dilemma kind of game repeatedly with the same people and eventually against everyone else (with a fixed strategy) to see which strategies were the most successful over all. Each player received a table or matrix representing what he will win or lose, depending upon his choice and his opponent's choice. The matrix ranks values as follows. Higher numbers represent greater value for 'me'.

	He cooperates	He defects
I cooperate	3	1
I defect	4	2

In this case we are shifting from a punishment style matrix as posited in the aforementioned prisoner's dilemma (that is, years in prison) to a reward style matrix where higher numbers represent greater value to "me".

In order to find strategies of greatest rationality, instead of people computer generated "agents" with specified strategies were the contestants. In the early runs, it turned out that the agent that always squealed (defected) didn't do that well, nor did the agent who always cooperated.

Instead the winning strategy was the one that did something called 'tit for tat'. Whatever your opponent did to you last round, you do to him this round.

Thus by this simulation, we find that a form of cooperation is a rational strategy even assuming self interest. This mechanism is supposed to explain many things, for instance, why gas stations on opposite corners aren't always in gas wars. They seem to agree to not undercut each other, even though they may have never talked about it or discussed it. Do these simulations in any way 'solve' the prisoner's dilemma? Not really. It does, however, explain why people do cooperate with each other in many cases and without contract to enforce cooperation.

Some try to draw moral inferences from the results of this simulation, some even suggest that it validates 'eye for an eye' justice. This I believe to be a premature conclusion. In particular cooperation is not a standard moral virtue. In fact there are laws that prohibit cooperation in some contexts, since sometimes cooperation leads to a virtual monopoly. It's curious to note that in the non-iterative prisoner's dilemma 'cooperation' is not the issue as much as who one is cooperating with (your buddy or the authorities).

This iterative prisoner's dilemma simulation may be the first philosophical uses of complexity theory. Complexity theory holds that when entities behave in parallel with some relatively simple rules of interaction what emerges is often complex and sometimes surprising behavior. Often entropy seems to break down, with order emerging from 'innocent' mechanisms. Complexity and simulations are said to explain many interesting things like how birds in flight, without a director, coordinate with each other and behave in unison. Or how a vast city like New York might have enough bagels for everyone every morning without a city director getting all the relevant information about need and supplies and planning it carefully.

One argument made in the book "Complexity" (M. Mitchell Waldrop, 1992) addressed the question of how life could arise on Earth when the probability that life organized by chance is so small. The calculations of the probability that a single protein molecule, which contains several hundred amino-acid building blocks in a precise order, would assemble itself by chance is so small that the probability that any earth like planet in the universe would ever bring forth life is still infinitesimal, even if you make the assumption that all of the trillions of stars in the millions of galaxies have earth like planets. By identifying the natural mechanisms that tend to organize, complexity theory has addressed this question, and in my opinion, successfully. (Although the Anthropic Cosmological Principle may be invoked to explain life, it does not explain the existence of life as much as it explains why we are aware of the existence of life. That is, it seems possible that life might never have existed, but it would be impossible for us to perceive that life never existed.)

But, we might wonder, is complexity theory a science? If it is science-like then its home is probably best in mathematics. Its major tool for exploration is simulation. In these simulations basic rules are posited and an initial state is defined. Then a computer calculates successive stages. Sometimes it is the process under observation and sometimes the final state is 'the answer' but in either case the study is something which in principle mathematicians could do. Because of the numerous calculations needed, a mathematician might need programming skills to finish the task practically.

Simulations, in general, are of two sorts. In one kind the rules are set with the intent being to imitate reality as accurately as possible, as with wind tunnels. The purpose of this kind of simulation is usually prediction, as in 'will the plane fly?' This is not the kind of simulation addressed by complexity theory.

Instead, a complexity theory simulation asks the question 'what follows from a set of rules about how each entity behaves with others when you put all these entities on a playing field?' The entities might all be playing by the same rules, or not. The playing field could be uniform, or not. The rules themselves may be stochastic, or not. Many times the researcher is in for some surprises. Sometimes some totally unpredicted phenomena emerges. It is here that we need the philosophically inclined to give interpretation to what has happened. One thing is clear. The simpler the set of rules, the easier the interpretation.

But now on to my thesis.

It was my original conjecture that ethics naturally emerges from advice giving and taking with those that we care about. It is not a surprise that people give practical advice to those that they care about as long as there is some expectation that the advice might be followed. For, if one believes that the advice is good and the person is likely to follow the advice then we would expect that the person be better off for having followed that advice and it should please us to find that someone we care about is better off.
But how does ethics come out of this? Let us suppose that there are two kinds of advice: practical advice and moral advice. Practical advice is of the tactical or strategic sort but moral advice is something else. Let me propose that moral advice meets these three conditions:

• 1) It says: Do x (or not x)
• 2) It implies: Avoid as much as possible and don't freely associate with people who do not do x (or do x respectively)
• 3) The advice would be given to anyone the advisee cares about.

This third condition is here to capture a sense of universalizability to the advice that many see as a necessary condition for morality.
There are a few reasons for wanting to give such advice. The major reason probably being that if the person follows the advice then they will be more like the kind of person the advisee would want to associate with. Another reason is that as a person who cares about the other, if he believes it to be good advice, he would believe that the other will be better off by following the advice.

My position then is that ethics derives from this kind of advice giving. We should note here that being moral advice is not meant to imply "correct" moral advice, or even good advice. In fact these may be different concepts. The standard for correctness must involve some other criteria. So how do we show that such an evolution is possible?

To show that it is not implausible that ethics emerges from moral advice giving, I decided to set up a simulation, the intent of which was to show that agents adapting to each other with relatively simple rules of interaction would naturally develop 'an ethic', or at least follow ethical advice. These agents would spread their behavior when it was successful behavior and die out when it was not. If in this harsh environment ethical advice-taking prevails under some circumstances, then we could conclude that indeed following ethical advice under those circumstances may very well be the way that ethics got its start and indicate that there continues to be naturalistic support for ethical behavior without need for institutions.

After many a complex beginning, incorporating degrees of caring, cooperation and cities of population concentration, I became convinced that I should keep things as simple as possible, at least at first . David Koskenmaki, who helped in the conceptualization of this project, had been urging simplification for some time, but only after a few crazy runs did it become clear to me how right he was. By simplifying we make the results much easier to interpret and we get the extra benefit that we can be more confident that there are no bugs in the program. And so let me present the experiment in its final stages.

To simplify, instead of dealing with all possible forms of ethical advice, I wanted to focus upon one kind of ethical advice. The advice: not to lie. If the simulations proved successful with this one kind of ethical advice, then I would argue that it puts the burden of proof on my opposition to show that other ethical advice does not or could not behave in the same way.

It was my hope then to show that such advice giving and taking emerges and maintains itself in a simple simulation of agents who get the chance to lie to each other, even in a world where lying is profitable and in which all have already developed sophisticated and successful Machiavellian tactical maneuvers. When parallel acting entities with simple programs change their behavior as a result of their experience, then we would call this a study in "Complex Adaptive Systems". This is the kind of experiment I wanted to do.

So, we let the agents in our simulation develop their strategies as they play against each other and on top of those strategies we add the ability to take advice, advice that sometimes goes against the strategies the agents have developed and are continuing to develop. It sounds unlikely at first that there would be any place for such advice taking, but in complexity theory there are often some surprises.

If there is lying then there must be something to lie about. To provide a context the 200 agents in the program walk around a 311 by 311 grid and when they are within 6 units distance of each other they play a little game that goes as follows:
One takes the role of potential liar, the other as believer. The potential liar, who we will now just call the liar, knows the values of the other. Those values are fixed in this way:

	Liar picks action 1	Liar picks action 2
believer picks action 1	4	0
believer picks action 2	1	3

This represents the payoff for the believer depending upon what each does.

At the same time the liar has one of 2 matrices that represent his payoffs
Matrix 1

	Liar picks action 1	Liar picks action 2
believer picks action 1	3	4
believer picks action 2	1	0

or

Matrix 2

	Liar picks action 1	Liar picks action 2
believer picks action 1	0	1
believer picks action 2	4	3

This is quite a departure from the prisoner's dilemma simulation since that matrix was the same for both players, and both players always had the same table or matrix to refer to. In reality, not every interaction we have with others pits our individual interest against another's. Sometimes when I do exactly what pleases me without consideration of what you might do and you behave similarly, we both get exactly what we want. And some consideration might be given to how the prisoner's dilemma interactions would work if a fuller context of all possible value matrices were included. And then of course we could also introduce the feature of not knowing what the other guy's matrix actually is. This would be interesting, but that's for another time.

Initially, I considered 24 different matrices for both the liar and believer, but realized that interpretation would be difficult without comparing it to something simpler. So instead in my simulation the liar has two possible matrices, the minimum number needed to allow the liar the ability to say two different things and be believed by a rational person both times. If we only had one matrix for the liar, like the first one for instance, then what would he say? If he picked action 1 and decide to tell the truth, he would tell the believer that he has picked action 1. However if he picks action 2 and lies then he would have to tell the believer that he picked action 1. The other two options (picking action 1 and lying or picking action 2 and telling the truth) have no foreseeable benefit. So for any rational liar he would always say: "I picked action 1." But then it seems that it goes without saying and the relaying of information has no important role. Ultimately this simulation may collapse to just another prisoner's dilemma simulation, but I didn't want it to do so immediately upon reflection. This is why I wanted to have at least 2 liar's matrices.

One of the two matrices is choosen at random by the program and given to the "liar" by the program at the beginning of the round. The believer does not know which of these two matrices the liar actually is given. The action proceeds in this way, the liar views his matrix and decides which action he will do. He also decides whether he is going to lie about which side he picked or tell the truth. The liar speaks. After hearing what the liar says, the believer then decides whether he will believe the liar or not. If he believes, then he will choose the action which would give him the highest personal outcome on the assumption that the liar is speaking truly. If he does not believe him then he will flip a coin and choose an action at random. We assume the coin is fair. Pr.(Liar picking action 1)=50%. The choices of both are then revealed and the payoffs are settled.

Let's take an example. Suppose the liar has matrix 1.
Possibility 1: If the liar picks action 1 and tells the truth and the believer believes then the believer will also pick action 1 then the liar will get 3 and the believer will get a 4.
Possibility 2: If the liar picks action 2 and tells the truth and the believer believes then the believer will also pick action 2 and the liar will get 0 and the believer will get 3
Possibility 3: If the liar picks action 1 and lies, saying that he will actually pick action 2 and the believer believes then the believer will also pick action 2 and both will do badly, 1 for each.
Possibility 4: If the liar picks action 2 and lies, saying that he will actually pick action 1 and the believer believes then the believer will pick action 1 and the liar will get 4 and the believer 0.

Thus, if we were just to play this game once and the liar were assured that the believer would believe then his highest outcome would be to pick action 2 and lie.

However, the liar never has that kind of assurance. So if he is not believed then he will get the expected value by chance for his decision. So if he picks action 1 and is not believed then he can expect to get 2 on the average. And if he picks action 2 and is not believed then he can expect to get 2 points on the average. To minimize luck as a deciding factor in the game, these are the values he will get that turn. We do not actually flip the coin but instead give the players the expected values.

If the liar has picked action 1 and the believer does not believe then the believer will get 2 .5 as this would be his average. If the liar picked action 2 and the believer doesn't believe then the believer will get 1.5 as the average of 0 and 3.

Both matrices are set so that the optimal outcome for the liar is found by picking the side with the highest value for him and lying and being believed. In fact, there is a certain dominance going on in his decision. The liar could reason as follows (if he has matrix 1): Either I will be believed or not. If I am believed then I should pick action 2 and lie, for then I would get 4 points. This is better than picking action 1 and telling the truth, which would only give me a 3. If I am not going to be believed then it doesn't matter what I do I will get 2, since that is the expected value on each side. So I may as well pick action 2 and lie.

With the first matrix we can look at this dominance in another way. Although with iteration there may be some surprises, without it there seems to be only 2 reasonable choices for the liar. To pick action 1 and tell the truth or pick action 2 and lie. In both cases he will be telling the believer that he is picking action 1. Let the probability that the believer believes him be a variable, call it Pr(B).
Then the liars expected value if he picks action 1 and tells the truth is Pr(B)*3+(1-Pr(B))*2 which simplifies to Pr(B)+2
The liar's expected value if he picks action 2 and lies is Pr(B)*4+(1-Pr(B))*2, which simplifies to 2*Pr(B)+2 . So he should lie since his expected value is Pr(B) higher when he lies than when he tells the truth. Thus, it would appear that no rational agent would ever tell the truth and so it would also follow that no rational agent would believe that the other was telling the truth.
However, imagine that if by an agent's past behavior he could alter the probability that the other believed him, although it may be unclear what behavior would do this, how much would the probability have to change in order for it to be worth a change in behavior? By letting the probabilities in the first equation and second equation be different probabilities we could solve for the break even point. So let Pr1 be the probability that the believer believes you if you always tell the truth, and let Pr2 be the probability that the believer believes you if you always lie. So we can find what Pr1 needs to be
Pr1(B)+2=2*Pr2(B)+2 which means that
Pr1(B)=2*Pr2(B)

So in order for it to possibly be worth it for a person not to lie, it must be a choice that at least doubles the probability of being believed in the future. So, if the probability of being believed is already 50% or higher when one lies, then there is no motivation on the part of the liar to be more honest. Notice that it is not true apriori that telling the truth more will make the opponent believe you more. One might image a strategy where when someone doesn't believe you, you then attempt to punish them by lying in the future so that they may become conditioned to believe you. However, a simpler view is that lying less will make you more believable, and this does seem to be the case, as we shall see later.

If we do the math for the believer, we see a simple relationship for the one time situation without iteration. If the probability that the liar is lying is greater than 50% then don't believe him, if less than 50% then believe him. If it is equal to 50% then it doesn't matter what you do. The believer also might wish to consider how he could change his behavior and make it more probable that the other be honest in the future. I'm not sure what equations would be interesting to take a look at here, except we may note that if you could make the liar tell the truth every time then you would get a payoff of 3 by believing when he spoke truly, and you only get a 2 if the liar lies all the time and you disbelieve. But in any case it is clear that the believer does not have a dominance matrix. His rational action will depend upon his assessment of what the probability of the other's lying is.

So this is what the agents are playing for and the point of the game from the agent's point of view is to accumulate the most points. The points will determine who dies and who reproduces. More on this later.

It's worth noting that it may not be clear what kind of behavior will make your opponent believe you more. Would it be better in some cases to lie more as a way of sending the message that he doesn't believe you enough and thus punish him, or is it better to lie less so that he will most likely believe you in the future? It may depend the nature of the opponents, or society, to determine which of these strategies is best.

What strategy will be the best in the long run? At this point I could have requested people to submit programs that play against each other in this arena to see what strategies were dominant. But there is a better way. If we could characterize all possible strategies and then pit them against each other, then this might do what we want. To do this we must characterize a strategy and as it turns out this is not too difficult. I'll discuss this after I describe the script in full.

So let me spell out the script of the game.

1: A round starts with the two hundred agents randomly walking around a 311^2 grid with a slight attraction to the center. Some of these agents take bigger steps than others.
2: We pick either matrix 1 or matrix 2 at random to give to the liar. (The same matrix goes to all to minimize chance elements)
3: We find all the pairs of agents within 6 units of each other.
4: Each takes the role of potential liar and potential believer in each of these pairs. So the first of the pair is the potential liar first and then the second of the pair gets the role, again to minimize chance factors.
5: Payoffs are made based upon their choosings.
6: Each individual's memory of what the other did, lied or didn't lie, believed or didn't believe, is updated. (Each agent remembers the last 4 encounters with each of the other 199 individuals)
7: An average gain is subtracted from all, so that if you had 3 encounters and the average gain for all in that round was 2 then 6 points would be subtracted from your total. This makes it so individuals who have more encounters per round do not automatically gain more than those who had 1 or 0. They have to do better each interaction than the population average to gain points over the others and over 0.

Then this is where it gets interesting.

8. After every 200,000 interactions the agent with the lowest number of points dies. His points are zeroed out and all memory of him and by him is erased. To replace him or her, the two agents with the largest number of points reproduce. They have a child that represents a synthesis of both of their strategies. The cost for this reproduction is that both agents points are lowered to the mean number of points (which is periodically set to 0). The population is thus maintained at the starting level which is set at 200.

By the way 200,000 interactions represents a life span of 65 years and an average of 8.4 interactions per day, (an arbitrary number, but I wanted a sense of time)

There are a few other arbitrary values and decisions that had to be made in this program, for instance after an agent has had 6 children, 10,000 points is subtracted from his total. This was done for the sake of genetic diversity. This almost always kills him off those two reproducers when the next two people die off. Almost every pair of agents are cousins after a few hundred years anyway with such a small population but I wanted to try to shake the genes up a bit.

The game continues either for a fixed period or until it is clear that a stability has been achieved.

An interesting thing to note is that these agents start out 'tabula raza', as dumb as rocks, with essentially random strategies, some believing those who have lied to them every time and some lying even though the other has never believed, and yet these creatures evolve through the mechanism described above quickly to some fairly intelligent strategies.

For instance, very early within the first 200 years, 98% or higher of the characters never pick possibility 2 or possibility 3 characterized above. Remember possibility 3 had no promise at all.
Possibility 3: He picks action 1 and lies, saying that he will actually pick action 2
There is no percentage in this strategy and our characters learn this really fast, even before advice is offered or taken.

But what is a strategy?
Each agent has a program representing what he or she will do with another based upon his or her experience with the other. So, for instance, one rule of a strategy for a believer might be...
Of the last 4 events with this person if when he said he would pick action 1 and he lied and then lied and then lied and then lied, then I will believe him next time he says he will pick action 1. Now this doesn't seem like a great rule, but someone probably has that rule at the beginning and that someone may die out really quick.
So the rule might be expressed as follows
If (he says he will pick action 1, he lied the first time, he lied the second time, he lied the third time, he lied the fourth time) then believe.

A strategy for believing is then all of those rules for believing that cover all possible situations.
Thus there will be 32 rules to cover all possible situations and 2^32 possible strategies, when you count a strategy as a set of rules. Each agent has exactly one of these possible strategies, from birth to death, which is to say that he has a fixed and complete set of rules all his life to deal with any possible situation.
Similarly there is a set for when the agent is in the role of the liar. A rule like that would have the matrix given and a history of whether he believed when that matrix was given. So an example of one of the rules might be
If (The liar is given matrix2, He was believed, believed, believed,~believed) then pick action1 and lie. (Where the "~" represents 'it is not the case that'). This time there are 32 rules but 4^32 possible sets of rules, since there are 4 possible consequent values for a liars rule. (He must decide which side to choose and whether to lie about it.) Now we have to have a few more rules to cover the cases where we haven't had 4 experiences with that agent yet, and so there are 62 rules for believing and 62 for lying. Thus, we have 2^62 * 4^62 possible strategies, or 9.808 *10^55

With numbers this size, you may wonder about the feasibility of finding a good strategy at all. That is where John Holland, the author of Hidden Order, and one of the original researchers in complexity theory comes in. He created an algorithm that zeros in very quickly on successful strategies, even when the number of possibilities is fantastic.

The algorithm is simple. Imagine as above that each agent has 124 rules and that these rules have an order or lay in a string. Holland thinks of each of the rules as governed by a gene. So the 124 rules could be represented by a string of genes. When two strategies mate, (somewhat successful strategies, remember) then the gene string is cut in an arbitrary place, say, for example, after the 30th rule, and the child receives his first 30 rules from one parent and the remaining 94 rules from the other.

This method of creating the newborn's set of rules is far superior to picking the rules at random from one parent or another and stringing them together, since rules can hang together. That is, for instance, if one parent picks action1 and tells the truth and the other picks action2 and lies then it might be horrible to let the child get one gene from one parent and the next from another for then he could easily end up picking action1 and lying, as we discovered. Holland's algorithm does minimum damage to strategies as a whole while sometimes creating a child whose strategies are better that either of the parent's strategies. Of course, sometimes the cut can still be in a poor spot.
Holland suggests the possibility of multiple cuts, but I decided to stick with one cut per mating, each time the cutting spot is chosen at random.

(I had initially set up traits to be passed on as a function of the trait gene from each parent. This involved making decisions like whether lying was a dominant or recessive gene. The Holland splice allows us to avoid that kind of decision.)

Also, so that the agents don't get stuck in a fixed pattern, each of the newborn's genes has 1 chance in 100 of mutating. For the genes that have two values, there is a 50% it takes on a particular value. Sometimes the new value is the same as the old and so there is really 1 chance in 200 that any gene changes. Without this all agents would likely become fixed on one strategy and it's likely that on different runs the final state would be different, depending on chance meetings and initial random condition. This mutation also interesting mirrors reality a bit.

The first runs then are done with just the 124 rules and no advice giving and taking. The run lasted 950,000 years or approximately 500 or 600 billion interactions. For this it took about a week of 24 hrs. a day computer time. Here is the results of one of three runs. About 99.5% of the time agents were associating with someone they had had more than 4 previous encounters with.

What we have in these three plottings is the percent of lying and believing (in each unit of time, approximately 30 years per unit) over 950,000 years in one run. What I find interesting about this is that even though everyone is behaving in a pure Machiavellian fashion it is still the case that in the first run 26% of the time lying is less than 1% (in a given time period), 38% of the time lying is less than 1.5%. The two other runs of a similar sort and in those we have 22.6 and 31.7% lying less than 1% of the time and 35.5 and 40.6% lying less than 1.5% of the time. Note that the believing and lying are moving together for the most part but in opposite directions. I think it is mostly believers responding defensively to the liars, but it works the other way too. The other thing that's interesting is that sometimes the society falls into some pretty stable states for 60-100 thousand years at a time. Looking for the explanation would be difficult. And why the stability suddenly becomes unstable again is also curious. This may be an example of what is commonly referred to as "punctuated equilibrium". Things were right for change with a little luck, and then things change over night, in evolutionary time.

With this graph I illustrate what percentage of the time lying is less than a certain percent. So for instance, we can see from the graph that roughly 30% of the time on each run, lying was around 1% or less. I think we can see here that the simulations behave quite similarily.

On some runs when liars were down at 1% and believers were greater than 99%, I decided to force the believes to believe at a 100% rate to see if the liars would respond. They did. It didn't take long for them to change their strategies so that they were lying at close to 100% of the time and it stayed at that level as long as the believers were forced to believe. It is as if the 1% that disbelieves is what keeps the liars in check. I also kept track of how many interactions were with agents will a full history (at least 4 experiences before.) That turned out to be about 99.5%.

Here is a graph where that happens when you force the believers into believing, and by the way it happened in all runs of this sort (I called this hypnotizing the believers).

This is a graph of one of those runs. It didn't take long for them to change their strategies so that they were lying at close to 100% of the time and it stayed at that level as long as the believers were forced to believe. It is as if the 1% of the believers that disbelieve is what keeps the liars in check. All of the runs eventually and rapidly converged on close to 100% lying and stayed there.

This is a shorter run where the believers were forced to believe around mark 5689. I suspect that the hovering around .5 for a while is because a liar learns how to lie about each of his matrices separately. Since there are two possible matrices for the liar, he has the possibility of having different strategies for each.

The run above was where each agent had a complicated somewhat Machiavellian strategy, each adjusting to the other. One might ask what the best strategy was? My best sense of it is that in depends upon the nature of the society as to what would be the best strategy at that time. At stability points in some runs, I decided to see what would happen if I were to drop in a character that I thought had a fine belief strategy but who never lied, a strategy that seems to work well about 26% of the time when the stability point is low in lying. Let's call him 'The Messiah'. Every time I tried it, manually, by stopping the program and changing the characteristics of someone to match The Messiah, The Messiah died out and never stuck around long enough to reproduce. Of course, if I were to drop him into the society when it was mostly like him, he would do fine. But if you drop this guy in when the society is not ready for him, he dies out. Also note that of all those times when people are lying less than 1% of the time, something eventually happens to change the dynamic. It is as if, everyone has found a way to be and someone very clever has found a way to take advantage of the situation. A new con. The con catches on and skepticism sets in leading to a less desirable state. By 'less desirable' I only mean that the total amount points gained on that round is not optimal. I called this measure of percentage against maximum possible points, a measure of societal health.

At this point, to provide some contrast like we might have with an experimental and control study in the sciences, I introduce an extra gene that has a propensity to accept advice not to ever lie. But because picking the wrong side and telling the truth would be a serious error, I took the advice to be complex.

• 1: If given matrix 1 pick action one and tell the truth
• 2: If given matrix 2 pick action two and tell the truth

If believed, both of these have less payoff than if we picked the opposite action and lied (and were believed.)

Because the characters start off as dumb as rocks, since the above strategy is much better than a random strategy, the characters quickly follow the advice, the society quickly falls into a state of lying less that 1/10 of one percent and a state of believing greater than 99.9 percent of the time. This was too easy and probably just reflects the fact that picking the 2nd best strategy of 4 strategies is better than picking a strategy at random.

So be more convincing about the value of advice-taking, we instead wait for a time when only 1.5% interactions are of the lying sort. We then introduce advice-giving and taking. To introduce advice-taking we insert an advice-taking gene. For the initial condition no one is an advice taker. Only by mutation (1 chance in 200) will this gene sometimes change to being an advice-taker. It would not have much of a chance of expanding unless it finds itself in an environment where advice taking is a plus survival trait and allows the advice-taker to reproduce at more than the average rate. Here was the results of one of the runs.

As the runs progressed, other questions seemed worth pursuing. Would hypnotizing believers work with advice-taking the way it did without advice taking? So after stable societies had established themselves and then after advice-taking was introduced and lead to "the permanent stability", agents whose role was to believe or not were hypnotized again to see if the liars would still be able to shrug off their programming and adapt and take advantage of the hypnotized believers. The answer is that in all cases (4 runs) they did find a convergence on close to 100% lying, the same as before. One interesting feature is that instead of it happening immediately, on one run it took over 100,000 years (one day and 1/2 of computer time) for them to find a way to make it work. The conclusion is then that the advice-taking gene provides more resistance to changing back to being good liars.

Since, my original thesis was that moral advice-giving was the seed of origin of morality and that part of that advice was to avoid or minimize contact with those who do not follow that advice, I modified the program to reflect that avoidance. So now, to adopt being totally honest also required minimizing contact with those who haven't adopted being totally honest. The runs were compared in the amount of time it took to converge on 1/10 of 1% lying (35 runs of each type). The result was that there was no significant difference in amount of time to convergence. Furthermore there was no hint that if the sample of runs were larger that there would be a significant difference. This does not help my original thesis and may count against it.

Because of a side thought, I became curious as to whether dividing the group of 200 agents into 2 arbitrary teams, the red and the blue, and expanding their programming so that they had the ability to respond differently to a member of their own team, versus a member of the opposite team, would cause the runs to have a different character. Would treatment of someone on one's own team be more charitable than with someone from the other team? I was looking for some sort of tribalistic feature. One might wonder why I would expect that. I really didn't but thought it would be interesting if it did emerge. It would be quite difficult to explain, and if it emerged it would provide many possibilities for future studies. What is the essential difference between someone on your team versus someone on another team? Well, with someone on your team, if his programming were the same as yours then he would treat you the same way you treat him. With someone on the opposite team, if his programming were the same as yours he would also treat you the same way you treat him. This is assuming that by same programming we are talking about genes that make the distinction between people on the same team versus people on a different team, not people on team A versus people on team B. So if you lie to everyone on the other team, and some member of the other team lies to everyone on the team opposite from his, then we would say that you have 'the same programming' in that case, even though you are lying to members of team B and he is lying to members of team A.

So why would there be a difference? Note that a member of the opposite team with the same programming as you, treats you the way you treat him, but treats members of your tribe differently than you do. It was my hope to find that this makes a difference in the runs. Unfortunately it didn't. There seemed to be no significant change in the amount of time it took to converge on final stability, nor was there any sense that agents were treating their own better that the others. So as interesting as it was, it did not work out.

(At the present time, after this last two paragraphs were written, I've discovered a problem with this part of the program. I had made the newborn a function of the strategies of the two top agents, whether or not they were from the same tribe. I assigned the new individual to the tribe of the guy who died and so this seems a mistake. I should instead find the two most successful people on the same team and mate them and make the child a member of their team. I can then easily test whether or not agents behave differently toward members of their own tribe. But when one team dies, I'll have to decide if I am to split the dominant team and start again or end the run. So there still may be a "tribalistic factor" after all. I'll be recoding and rerunning this part.)

Other runs were tried with various memory lengths, a memory of 4 long seemed reasonable since the ones that were seven long seemed to behave the same way but were much slower because there were so many more genes to evolve. Memory lengths of 1 seemed too far from natural experience to be interesting. Some runs were done where the order of the last set of events was treated as irrelevant, so instead of considering that he lied, he lied, he lied and then he told the truth, the agent would only note that he lied 3 times out of the last 4 occurrences. This also didn't seem to matter in the initial runs and so I arbitrarily chose to keep memory of order.

Because of some concern that the lying/believing simulation might just be a variation on the prisoner's dilemma cooperation/defection simulation, I constructed a few runs where agents had the option of lying, telling the truth or saying nothing. It could be argued that 'saying nothing' is more like defection, where truth telling is cooperation and lying is something worse than defection, like subversion. To see how things might go I made it so saying nothing was equivalent to both sides getting their expected value. There was no reason to believe that the behavior of saying nothing would actually be a winning strategy, but because unexpected things happen with complex systems, I ran it anyway and the result was as expected. Those who refrained from speaking did not do well and died out quickly. After that result (with several runs), all subsequent runs were done without the 'saying nothing' option.

While writing the end of this it occurred to me that another interesting prisoner's dilemma simulation could be done with no memory of the past but with the knowledge of whether or not the other has programming like your own. That is, when identical twins are separated and grilled, with the assumption that your twin will likely do what you do, you may end up with the Newcomb type paradox, which by the way might also be worth exploring with a simulation also, except that wouldn't be adaptive because of interaction.

Since I could make small modifications in my program and accommodate this question, I ran it and it came out with the following predictable result.

With this run I started the strategy I thought would win with the handicap of having no initial representatives (cooperates with twin only). Only by mutation and thereafter natural selection would this strategy emerge as a clear winner. It did. Cooperate with those who are like you and defect against the others, was the winning strategy

It is clear, at least in my ideal society, that at least one moral rule emerges and is sustained without institutions being involved. In fact it would be an interesting project to see how we would even define and implement "institution" in a simulated context. I suspect that a definition is possible but without intentionally being inserted into my simulation, it would be quite a surprise if institutions emerged. So the major thesis, that morality could be created and maintained without institutional support, is well confirmed.

But we may ask why it is that these simple advice genes (the "never lie" genes) become more powerful and the dominant force over carefully constructed and complex strategies? Some have argued that the reason 'not to lie' and 'to be moral' in general is that because telling the truth and being moral is easier to keep track of. You don't have to remember who you lied to and since we are all fallible we won't get caught at anything that involves a punishment. But my agents have perfect memories and all agents are told at the end of each interaction whether the other lied or not. They remember who they lied to and how much the others believed them. So it is not the agent's fallibility that is a major reason for the domination of the most general and simple rules of behavior. I'm at a loss to explain why simpler rules turn out to be more powerful forces toward a stable and healthy society. It's one of those unexpected results that sometimes happens with Complex Adaptive Systems.

One objection to my argument might be that there are many structures going on in the simulation that are disanalogous to real society. After all, people don't behave by a program of action without reflection and clearly our society has not converged on a place where 1/10 of 1% of utterances are lies. In our simulation, agents find out immediately whether they were lied to or not. This is also not analogous to real life either. All of these 'reality' features are a problem for a solid conclusion, but if we take the agents strategy gene as not so deterministic in reality but as a tendency towards strategies of a certain sort and take it to be the case that a few lies are discovered to be lies in reality, we might be able to explain why more than 1/10 of 1% of utterances in reality are lies.

Furthermore, these agents do not learn, like we learn. When we try something and it is unsuccessful, a reflective persons ask 'what was my mistake?' 'What change in behavior would make a difference?'. It would be interesting to make it so that not only did agents die off when their strategies were weak but that they also changed their strategies if they proved to be unsuccessful. This is interesting but when I think about programming this, the result is merely a duplication of the die off and reproduction idea of how to change. Note that when a strategy fails, it might be difficult to infer what strategy would be successful except to "not do that anymore". Reasoning can be helpful at the beginning when people are "tabula raza" but when everyone is adjusting to everyone else, who happen to be adjusting, reasoning is difficult. I would suggest that evolution is the best we can come up for finding the answer to "what is the most successful strategy at any time of the state of that society?".

But instead of treating the program as a model of reality or real societies, we can instead think of it as exploration into what the rational course of action would be for these players in the lying/believing game context. Let us not think of the simulation as a model of real people dying and being born with certain strategies, but instead as an attempt to model learning as strategies die out from lack of success and are replaced with better strategies. As such, we would interpret the result as confirming at least with respect to lying and believing that there is a natural mechanism that eventually leads to a stability point of very little lying. If we characterize this as a force then we have essentially answered the opponent argument, if his argument is that without institutions people would have no reason not to lie, cheat, and steal. We have found a mechanism or force, at least with respect to lying, that predicts that people would not lie all the time, even if there were no institutions. The burden of proof then moves to the opponent to show that there is no such mechanism in real society. Furthermore, since this mechanism was found with respect to lying, there seems no reason to believe that there are not analogus mechanisms for the rest of morality. To counter this the opponent would have to provide a reason for why this would not be the case.

So instead of being able to come to the conclusion that 'this is the way things would have developed without institutions' we instead have provided reasons for why things might have gone that way, and likely would have as long as there was no other forces to counter it. In the final stable states of the runs with advice-taking, we find that those who are not moral fail rather quickly. At least in that context, to be moral is rational. It seems likely that if self-respect and being the kind of person you would wish to associate with was also a feature of the program, that this would count even more strongly as a moral stabilizing force.

As for my original thesis that ethics emerges from giving advice to those you care about, I think we have to say that although it is compatible with the results it is not entailed by the results. For one thing, the program actually lacked the feature of coming in contact with an advice giver. It assumed that the advice was 'out there', since reasons were given for why people would naturally give advice and I preferred not to add that level of complexity to the program. I suspected that it would slow things down. But since everyone has many interactions with everyone else, if contact were required to pass this advice, it should not look much different from what we have. Any single person giving the advice would give it to everyone fairly quickly. The second problem for my thesis is that avoidance of the non-moral agents did not seem to play a role in the result. I would have thought that that part of the advice would be a strong force, instead it is no force at all.

Another series of questions could be asked. What about that first guy who decided to take the advice to not lie? Given that many had tried before and all unsuccessfully and with no reason to believe that his actions would prevail and become the dominant paradigm, would we still say that we was doing what was right? When you are in the extreme minority, does the probability that your message through actions be heard influence the rightness of the action? How about the ones before him who tried it and failed? And what about actions that would clearly lead to a better world if everyone would follow them but have no chance of generally catching on? Are those right actions? As Mr. Smith who went to Washington said "Maybe lost causes are the ones most worth fighting for". (What he actually said was that "Maybe lost causes are the only ones worth fighting for", but that can't be right.)

One curious thing happens after the advice gene takes over. The less dominant Machiavellian genes underneath become irrelevant and can take on any uniformity, by chance. When they no longer have a strong influence, random mutation and a successful grandmother can make the whole society have the same irrelevant set of genes. If the advice gene were then removed, a chaotic search for balance would begin again. This also might explain why when believers are hypnotized when there are advice genes, it takes a long time for the liars to take advantage of it. After shedding the advice-taking gene they also have to rediscover how to lie well because the underlying strategy genes have been randomly evolving without a success selection mechanism of any worth.

Could it be that lying and believing is just a variation of cooperation and defection? More specifically lying might be like defection and telling the truth might be like cooperating, whereas not believing could be like defection and believing would have the same role as cooperating. The answer to this would be interesting no matter which way it turns out. There are some clear differences in the simulations herein and the prisoner's dilemma simulations, namely the believer anyway doesn't have a dominance matrix for his decision. What he should do for one time personal gain is dependent upon his assessment as to the likelihood that the liar is actually lying to him. In the prisoner's dilemma, both players are forced into not cooperating no matter what the other does, if all they are concerned about is their own one time personal gain. This makes the runs clearly different, but it could be argued that the prisoner's dilemma does not represent all cases of cooperation and defection.

The issue of whether my simulations and the prisoner's dilemma simulations are examples of more general cooperation/defection issues will depend upon how we define cooperation/defection. I would think that one is cooperating with another if one does what he knows the other would want him to do. We know that the liar, given that he doesn't pick a dumb strategy, always wants to be believed. The believer always wants to be told the truth, except in those cases where he is not going to believe, in which case he doesn't really care what the liar does. With these similarities and differences we can see that this argument could go on in an interesting way.

Can we decide now that it is rational to be moral? It takes a bit more argument. If the sim agents were fleshed out a bit we could offer the following argument. The sim characters have programming that leads to following moral advice. Whenever one does what one has the innate desire to do then one gets some satisfaction from that action. If one gets satisfaction from an action then that provides an extra reason for doing that action. So because the sim agents have genes that direct them to follow moral advice, this gives those characters some purely self interested reasons for following moral advice, just because it is moral advice. And thus following moral advice is going to be rational in more cases than would have been if we neglected these satisfactions. If we are anything like those characters, we would get satisfactions from following moral advice also. This is not a solid argument for always doing the right thing but it does add weight to the right-thing side of the equation.

We have focused so far only upon moral advice and not correct moral advice. What would correct moral advice look like? Certainly we could make morality being derived from moral advice compatible with any self-consistent moral theory, but the simulations might lead us to focus upon a couple of interesting features. First of all, in our example when the characters by and large follow the moral advice the overall health of the society (measure as collective payoff) is maximized. Secondly we might note that some moral advice has a strong chance of universally catching on whereas other moral advice might never catch on. It would be tempting but definitely premature to conjecture then that correct moral advice would be advice that could lead to an optimal state if everyone were to follow it and furthermore it is possible that the advice catch on and lead to a stability in that optimal state (maybe inevitable, if enough people in the minority try from time to time to follow that advice.) Just a thought.

I suspect there are many ways to address other philosophical and social questions using complexity theory. Instead of asking how did things evolve or what is the rational course of action, with complexity theory we can address other equally interesting questions: "How might things have evolved?", "What could be the rational course of action if we were beings of a specified sort?". What would we call this kind of study? Surrealistic Philosophy or Sociology? That might be an amusing way to think of it.

Thank you for your interest and patience.

The following reveal ideas and earlier stages of the experiment.

• page1
• page2
• page3
• page4
• page5
• page6

Thanks to David Koskenmaki, Gary Shannon and Don Salter for their important contributions in the dialogues that we had in trying to conceptualize the project that seemed at first like nailing Jello to a tree.