Computer Simulation of Multilevel Selection and Evolution

The topic of "Multilevel Selection" has been going around lately, in the context of how biology and the genetic selection that Darwin described, meshes with culture and religion and other "meme" selection that has been advocated by Richard Dawkins and his advocates. A similar concept has been put forward by E. O. Wilson, who calls it "sociogenetic" selection, to indicate that the culture as a whole works together with the genetic material to produce some sort of emergent collective organism (Wilson looked at ants which would have to be the most startling example of a highly organized social structure operating almost like a single entity, see also "Sociobiology").

Well, there has been some disagreement between Dawkins and Wilson over the operational details of this mechanism, and on top of that, David Sloan Wilson is also buying into the discussion and he often calls it "group selection", but as far as I can tell, David Sloan Wilson uses "multilevel selection" and "group selection" quite synonymously. Here is a recent video on this topic:

EIGHT CRITICISMS NOT TO MAKE ABOUT GROUP SELECTION

Major topics in science typically include a zone of consensus among the experts, based on past progress, and a zone of current controversy still being thrashed out. The topic of group selection is unusual in the degree of ignorance among professional evolutionists who are experts on other topics, but whose ideas about group selection remain sadly out of date. Their ignorance interferes with the peer review process when they are called upon to review articles on their area of expertise that are written from a multilevel perspective. David Sloan Wilson and Omar Eldakar review eight criticisms that no one should be making about group selection, based on a commentary published in the academic journal Evolution in 2011.

My Take on the Selection Theory

I've come to the conclusion that evolution must happen on every level simultaneously because there is no way to justify a privileged position for any particular evolutionary unit. Of course, in some cases, picking a particular unit of reproduction to base your theory on can be an excellent simplification, and every theory requires suitable choice of simplification, but just remember that it is nothing more than a convenience.

I'm certainly not a biological specialist, so I dislike bilologists using words like "altruism" and "group selection" to mean something special to biologists, but just about every non-biologist would take them as meaning something else. The constant confusion between concepts such as "altruism", and "reciprocal altruism" (which should more correctly simply be known as "trade") tends to spring from people getting anchored into a particular way of looking at things. That is to say, they focus entirely on the individual and completely ignore group effects, or focus entirely on the group and ignore the individuals.

In particular, "group selection" is a terrible term to use and should be completely avoided. I'm sure David Sloan Wilson understands what he means by it, but the impression any normal person would get is that somehow selection amongst individuals is no longer happening, but of course selection is always happening at both levels -- amongst the individuals, and also at the group level as well. The term "multilevel selection" conveys this concept clearly, making it the preferable term.

Computer Simulation of Biological Problems

Many practitioners of the physical sciences have an uneasy relationship with computer simulation -- and rightly it should be uneasy. Simulation should be treated with suspicion because essentially a computer can be programmed to do anything you like, doesn't mean it every happens out there in the real world. Thus, often the results of a simulation do nothing more than demonstrate the predjudice of whoever designed and programmed that software.

The thing about evolutionary biology is nearly all of it is observation, and drawing conclusions based on extrapolating the observed case. Here is the answer, I'll make up a question that seems to lead toward that answer. The evolutionary biologists are a bit weak when it comes to experimentation and predicting the results of their experiments. Laymen think that studying evolution ends up nothing more than a collection of "just so stories"; great as a comforting explanation for what you already know, but useless for producing genuine new discoveries that you didn't already know. Once again, the "just so story" method of retrospective explanation often tends to reveal more about the predjudice of the person explaining, than it does about the system itself.

Computer simulation provides some sort of workable and safe option for putting evolutionary theories to the test. Like all tools it must be used appropriately. Let's suppose you want to simulate a real physical system, like you want to discover exactly what the average global temperature of the Earth will be in 100 years as a result of additional CO2. Those sort of simulations require incredible attention to detail, and careful callibration, especially if you regard the difference between 1 degree warming and 2 degrees to be highly significant. The simulation presented here is not even an attempt to model any physical system, it is only a demonstration that the basic mechanism being discussed can work. There has been no attempt to callibrate this or to model any particular cultural or religious difference here. It is a very simple and generic model, designed to be a minimalist starting point to encourage experimentation.

Some of the arguments levelled against the "multilevel selection" theory are along the lines of, "It is illogical, and it simply cannot work at all."

Hopefully, this simulation does at least prove, yes the mechanism can work, there is no particular mathematical rule that prevents it, because it has been demonstrated. That does not prove that exactly this mechanism is at work in any real life scenario that you care to name; however since the result comes out as an emergent property of a very simple model, and since all of the assumptions that go into it are quite reasonable and plausible, we might expect to find some similar mechanism at work in the real world.

Fundamental Components and Assumptions of the Simulation

• The world model used is a grid of "cells" similar to a cellular automaton, such as Conroy's "Game of Life", each cell represents one individual entity. In these examples, the world grid is chosen to be 640 x 480, for display purposes, so each individual can be represented easily by a single pixel on an image.

• The individuals contain certain parameters and interact with other nearby individuals in a small local area around themselves.

• Interaction between individuals is updated on a uniform random basis, so that over many such interactions, the world grid progresses. Each interaction also introduces tiny random variation to the parameters that define the individual, allowing evolution to occur.

• One of the values recorded against each individual is a measure of the resources owned by that individual (we consider all resources to be just a homogeneous quantity). Individuals are considered "dead" if their resource value drops to zero, and the empty space can be repopulated by other nearby individuals reproducing themselves. Every time an individual is updated, some fixed amount of resource is consumed as regular upkeep, so individuals must collect more in order to avoid death.

• When individuals reproduce, they copy themselves exactly, with all parameters being identical except that they total resource value is split in half, with each copy taking their own half. Thus, reproduction neither creates nor destroys total resources.

• Each interaction between individuals is modeled as a "Prisoner's Dilemma" situation, the elaborate details of this have been split out into their own section, but in brief, the important points are: this a non-zero sum game; players in this game may choose their approach on a continuous scale from completely uncooperative (a value of 0.0) to completely cooperative (a value of 1.0); only cooperative interactions between individuals can deliver a nett positive and thus overcome the overall upkeep requirement allowing survival to be possible at all; however there is nothing preventing some individuals from exploiting others (at a resultant cost to the groups as a whole).

• Another of the values recorded against each individual is a "Cultural Identity", which is represented as a 32 bit Hamming space. This is a bit of an unusual approach and has been dealt with in detail in its own section. In summary, individuals can detect how similar/different they are from their neighbours and use this information to decide how cooperative or uncooperative they might be in this particular interaction. Most importantly, individuals have the power to learn and will choose to make their own cultural identity more similar to those who have treated them with kindness and less similar to those who have treated them harshly. This is the strongest assumption, because the learning mechanism is not an evolved trait -- it is hard coded. However, it is not entirely an unreasonable assumption.

• Each individual keeps a parametric tolerance function which provides a mapping from the cultural Hamming distance between two individuals onto the individuals attitude toward the "Prisoner's Dilemma" game. In other words, this decides who the individual will be cooperative towards and who they will be hostile towards, based on cultural idendity. Note that the parameters of this function are free to evolve, but the inputs and outputs are clipped to the range 0.0 to 1.0 on the basis that these represent the most extreme values possible.

• The evolutionary process continues as some individuals are eliminated and other fill in the space with copies of themselves.

• Every so often an output frame is generated (in Netpbm P6 format) using an approximate mapping of cultural identity bits onto colour of the pixel and total resources owned by that individual becomes brightness of the pixel (dead cells are black, i.e. no colour and no brightness).

• Log data is output to the stdout, containing samples of interactions. This allows statistical analysis of the process.

• ga_2012-07-02.c -- Genetic algorithm simulation program.
• ga_2012-07-20.c -- Genetic algorithm simulation program (small changes).
• gimp_proc.pl -- Perl script that uses the "gimp" program to blur each frame and save under a different name.
• pnm_YUV4MPEG2.c -- Video frame converter for movie output (PNM to YUV).
• process_frames.sh -- Script to assemble frames into movie (OGG format).

If you want to make your own movie, the three steps involved are driven by the scripts above, and you will need to install "gimp" and "theora_encode" which are commonly available on most linux distributions. TODO: the GIMP Animation Package (GAP) might be able to output to ogg/theora and do this in one single step, or I should write a program that can do it in one step.