Fabio Ruini’s blog

A trend that is both disturbing and interdisciplinary emerged during the 1990s and is continuing to expand at an alarming rate. This trend involves the creation of really bad evolutionary models by graduate students (and some professors) in a variety of disciplines, including computer science, biology, cognitive science, and physics (cf. Shalizi & Tozier, 1999). We examine the trend and derive an ignorance-based algorithm for obtaining a Ph.D. that is robust across many disciplines and can adapt to such volatile conditions as lack of funding and absentee advisors.

Artificial Life (A-Life), according to one widely cited definition, is “The study of man-made systems that exhibit behaviors characteristic of natural living systems” (Langton, 1989). For many legitimate biologists, computer scientists, and other researchers who defy classification, this new field offered the prospects of collaboration across disciplines and new methodologies for investigating vexing problems.

Alas, however, A-Life has become a safe haven for would-be scientists that are unable to face reality but sometimes enjoy reading about it. We do not wish to implicate specific persons in our critique, so we will follow the pioneering paradigm of Shalizi & Tozier (1999) and extract an overarching algorithm that seems to be in use among many graduate students carrying out A-Life research.

Shalizi & Tozier (1999) focused their critique on the behavior of physicists. We aim for a wider, more interdisciplinary target. The algorithm for a Ph.D. is as follows:

1. An undergraduate student becomes interested in Artificial Life through reading pop-science A-Life books such as Levy (1992) or Dyson (1997).
2. He decides upon a career in A-Life, and majors in a relevant field, such as computer science or biology, or a marginally relevant field, such as cognitive science, chemistry, or mathematics, or an irrelevant field, such as French. The actual diploma that he receives is not of consequence, as long as he has a basic understanding of one programming language.
3. He is accepted to some doctoral program where either of the following are true:
   • Artificial Life research is being perpetrated by at least one professor, or
   • He can be left alone to work on whatever he pleases (the absentee advisor problem/benefit).
4. He is encouraged to read A-Life journals, such as the journal Artificial Life. He is not encouraged to read anything outside of the A-Life literature, such as journals/books from the various disciplines that comprise A-Life. For example, by the time this algorithm is completed, he has not picked up a single copy of Cell or ACM Transactions (on anything) or Nature or Science. He thus runs into the same problem, which we will call disciplinary inbreeding, described by Shalizi & Tozier (1999) for the case of physicists.
5. He picks a topic, usually within the domain of evolutionary biology, that is of interest, and enters subroutine 1.
6. Subroutine 1 is called:
   • He develops a “theory” that can be “satisfied” by “results” of an A-Life simulation.
   • He writes a multi-agent simulation, probably based closely upon one he has read about. Due to the great number of parameters available for setting, he calls this simulation a “model.” For good measure, he adds some new parameters, which are generally derived from his imagination. Also, the simulation must generate pretty pictures.
   • He devises a set of measures that, when applied to the simulation after an appropriate amount of time, will show the “validity” of his theory.
   • He presses the Enter key and goes to lunch. Because of the inefficiency of his algorithms, use of improper data structures, and inappropriate choice of programming language, he will return the next day to check on a simulation that could have been completed in eight minutes.
   • Returning the next day, he finds that his measures are not as expected. At this point he applies a genetic algorithm (GA) to his simulation. The inputs are the parameters of his model, and the fitness function is based upon the goodness-of-fit to the expected values of the measures that will boost the case for his pet theory. He presses Enter and leaves.
   • Returning a month later (we skip the nine iterations of memory leaks and segmentation faults he has fixed), he finds a reasonable fit of the results to his expectations. He collects the values of the parameters that produced the good results.
7. Subroutine 1 ends. The values of the parameters and measures are returned.
8. Relying on intuition, imagination, and a keen ability for storytelling, he creates a justification for the value of each of the parameters discovered by the GA. For example, “We set the mutation rate to 28% of the genome per generation because, during the Triassic period, numerous volcanoes erupted.” Everything after the ‘because’ part of this statement should be factual, but need not be relevant to what came before it.
9. He writes a paper and submits it to one of numerous A-Life journals.
10. He repeats this process for several slight variations on his topic of interest, submitting a slightly different paper each time. The differences will most often come from radically different sets of parameter values discovered by his system under conditions in which one or more of his measures is perturbed by a small amount. Again, his intuition, imagination, and a keen ability for storytelling lead to a new justification for these parameter values. He does not realize that, were he to run his simulation more than once while feeding the GA the same expected values for his measures, it would produce a different set of parameter values each time.
11. Peer reviewers do not realize this either, as they are busy eating lunch. They skim the paper to make sure it contains such keywords as “evolutionary”, “dynamical system”, and “volcanoes”. The paper is published.
12. After several iterations, he is awarded the Ph.D. on the justification that he has several peer-reviewed publications, even though no one can understand what the hell he is talking about at his doctoral defense. Many audience members are, however, very taken with the movement of various colored dots around the screen, and the interactions between those dots. Whether the dots represent genes, organisms, populations, or something else, is not readily apparent.
13. A post-doctoral appointment may follow, wherein the algorithm is repeated twice.

Langton, C.G. (1989) “Artificial life.” In C.G. Langton, ed. Artificial Life, Vol. IV of SFI Studies in the Sciences of Complexity, pp. 1-47. Redwood City, CA: Addison-Wesley.

Levy, S. (1992) Artificial life: the quest for a new creation. New York: Pantheon Books.