Digital Digital Burgess Conference Follow-up:
Conference was held August 29-September 1 1997, Banff Alberta, Canada.
Artificial Life and the Natural Sciences
Bill Riedel (et al) -
This is in response to your request for short descriptions of the applicability of our various research areas to the natural sciences and vice versa. I hope the document you envision comes together and serves as a valuable reference for everyone, both as a source of insight and inspiration, and, perhaps, as practical support for various grant applications.
There are a couple of general, introductory paragraphs first, and then a few paragraphs specifically on PolyWorld's relation to the natural sciences. I hope this fits what you had in mind. I wasn't sure what voice to take, and so referred to myself in the 3rd person, though you may want to change that to first person and stick my name on the section. Feel free to edit this for the sake of consistency, length, or whatever. If other ALife people also write general introductory paragraphs, you may want to rename this "PolyWorld and the Natural Sciences". If other authors' writings are specific to their own work, you may want to lump us together under this heading. Whatever works out best. If you want me to review changes to this, I'll be glad to do so; please just email the revised copy to me.
Artificial Life and the Natural Sciences
While the fields of Artificial Intelligence (AI) and Artificial Life (AL) may have some grand goals in their own right--including true machine intelligence--they are also profoundly influenced by the natural sciences, and may offer some valuable insights to the natural sciences in return. Indeed, the transition of focus from AI to AL is primarily inspired by evolutionary and biological insights from the natural world. Specifically, while AI attempts to directly model the high-order phenomenon of human-level intelligence, AL attempts to work bottom-up--modeling better understood, low-level phenomena and seeking to obtain higher orders of complexity as an emergent property of the interaction of this substrate. AL also naturally lends itself to a literally evolutionary progression of behaviors from the simplest to the most complex adaptive behaviors--from primitive individual survival mechanisms such as one might find in a nematode to a complex social mix of self-aware intelligences such as one finds in human society.
Since AL borrows heavily from observations and understandings derived from the natural sciences, it may also provide some useful feedback to those studies. And though AL models may sometimes deliberately diverge from more natural systems, such models may also derive significant benefit from maintaining their close ties with the natural sciences. Just as a computational fluid dynamics model is benchmarked against wind tunnel test results, so might a computational ecology be benchmarked against natural ecological data. Both agreements and discrepancies can be of significant interest--to both the AL modeler and the behavioral ecologist or the evolutionary biologist. The incredible variety of life discovered in the Burgess Shale, that has documented and demonstrated the Cambrian explosion of life, owes that vast diversification, along with the specific traits to be found in those (and all) living organisms, to evolutionary and ecological processes that we have a chance to explore and better understand through our AL computational models.
To highlight a few specific instances where AL might feed back into the natural sciences, consider Larry Yaeger's PolyWorld simulator. There are a number of specific simulation experiments that might provide some useful insights into natural evolutionary and ecological processes.
One such experiment, that relates to the interests of Evolutionary Biology, might simply explore the relationship between speciation and geographical isolation of populations of organisms. Imagine tracking the distribution of genetic disparity in and between several populations of organisms in a series of different simulations permitting varying degrees of contact between those populations. So in one extreme the populations are completely isolated. At the other extreme the populations intermix freely. The "barriers" of PolyWorld easily permit such isolation (or lack thereof). In between the extremes it is also straightforward to run simulations that block 95%, 90%, 80% (and so on) of the border between these populations. So at least for one set of initial conditions, then varying only this degree of isolation, one might obtain an approximately continuous measure of speciation as a function of population isolation. Perhaps suitably semi-isolated real geographic ecological niches could provide natural data to compare and contrast with such results. No code changes to PolyWorld would be required to explore this subject, except perhaps a few lines to record the specific, desired measure of genetic diversity (over time) to a file.
Another such experiment, also rooted firmly in the Evolutionary Biology world, might be an investigation of the emergence of altruism. Though altruism is not built into PolyWorld, it might be an emergent phenomenon one could study there. Again with no substantive code changes, one could extract a distribution of the degree of predation practiced by all the world's organisms as a function of genetic (dis)similarity. If predation were found to increase as a function of genetic dissimilarity, one could reasonably claim to have discovered altruism in this computational ecology. With some slightly more substantive code changes, one might be able to explore the effect of the availability and growth patterns of food, or other ecological pressures, on the degree of altruism.
Dr. Larry Dill, a Behavioral Ecologist at Simon Fraser University, suggested that a minor variation to PolyWorld might permit a fairly rigorous modeling of a crucial aspect of Orca whale behavior off of Van Couver Island. The whale population there has split into two distinct groups, one of which feeds from an essentially fixed food source found on an island (seals), and the other of which roams more widely and consumes an essentially mobile food source (fish). The complete Orca population of only about 300 whales could be effectively modeled, with some portion of PolyWorld's food being forced to grow in a fixed location, and the remaining portion of that food being mobile. It would be interesting to see if PolyWorld (or "OrcaWorld" as this variant was dubbed) would evolve the two distinct populations. It might also be possible to investigate relative population sizes in the two groups given varying proportions of the two sources of food. Some historical data about relative food distribution and pod sizes might permit some interesting comparisons with such a series of simulations.
The final such PolyWorld experiment that will be mentioned here is one that also falls within the purview of the field of Behavioral Ecology, and relates to optimal food foraging strategies. In particular, Larry Yaeger designed a sufficiently simple foraging vs. predation model, in a variant of PolyWorld, that a formal analytical solution is possible to determine the optimal foraging behavior. Imagine a world divided in two. On one side of this world, food grows freely, but there is a small probability of an organism being killed with each unit of time (a time step in the simulator). On the other side of the world there is no food at all, but neither is there any predation risk. The value of that risk of being killed, the predation risk factor, then decides what proportion of an organism's life should be spent foraging, versus loitering safely on the risk-free (but food-free) side of the world. This optimal ratio of time spent one one side versus the other can be determined analytically. One could run a series of simulations to see if PolyWorld's organisms evolve this optimal solution for various different settings of the predation risk factor. Both the analytical solution and, of course, the simulation technique differ somewhat from the traditional Marginal Value Theorem approach to the analysis of foraging strategies. Those differences and similarities should be interesting and revealing.
Certainly other crossovers between the natural sciences and PolyWorld are possible, and some basic lessons have already been learned about free energy, energy transfer mechanisms, and the availability of simple, non-anthropocentric explanations for seemingly complex, ethological behaviors. And other ALife systems have yielded their own insights. Tom Ray's Tierra simulator has demonstrated the evolution of parasitism and hyperparasitism. Karl Sims's evolving creatures have demonstrated an exciting convergent evolution between natural and digital organisms for certain walking and swimming behaviors, and demonstrated how competitive forces help drive speciation. ALife simulators give us the opportunity to explore many different biota, and relate them to the natural biota in which we live. Or to paraphrase Chris Langton, ALife is the study of life as it might be, so that we might better understand life as it is.
by Larry Yaeger