|
Questions and Answers During a meeting (1st October 2004) with people from theoretical biology and mathematical biology, several question were posed and suggestion were made about the methodology and projects of the Virtual life lab. We hope to clarify some of the issues raised. Q: Why pitch the simulation on such a high level? Q: Why do research in theoretical biology, and not stick to evolutionary robotics? Q: Genetics not important for evolution ?!?
A: Simpler simulations are very effective, but only on the level of complexity they deal with. Game-theoretical simulations have revealed some fascinating mechanisms involved in the evolution of cooperation (Tit-for-Tat in the Prisoners dilemma; Axelrod, Hamilton). However, such simulations readily presume rational decision making. This may not be a sound assumption, since cooperation exists in many organisms that are not capable of making rational decision (e.g. ants). In other simulations, individual-based with grid-world environments, agents are able to execute one of several behaviours. For example, move North, move West, etc. Because these simulations are specified (modeled) at the level of rule-based individuals, these kinds of simulations can show interesting spatial behavioural patterns on the level of populations. Our interest, however, is in the evolution of behavioural patterns on the level of the individual. The individual organism interacts with its environment based on its sensor-motor coordinations. To investigate the behaviour on the level of the individual, one needs to specify entities on a lower level. That is, one needs to model the body, its environment, sensors, muscles, and the interaction processes between these components. Why pitch the simulation on such a high level?
A: The answer to this question is related to the one above. This question is nonetheless interesting, because it reveals an interesting misconception about other simulations. Framsticks simulations, with all its complifications, admittedly has many implicit assumptions. These assumptions are, however, pitched on a very low level (neuronal activity, inertia, gravity). The assumptions in other kinds of simulation are perhaps less in number, but are much more high-level. Game-theortical simulations, for example, have just one assumption, but one that is extremely improtant and high-level: the entities in the simulations are capable of rational behaviour. Presented as such, Virtual life simulation are actually pitch on a rather low level compared to other kinds of simulations. Although it is possible to model even lower levels (quarks, proteins, cells, etc.), one can hardly expect to gain understanding about behaviour of the individual from this.
A: Nothing! Nothing is inherently new about the results themselves. What is new, is the way of obtaining the results. This is exactly the purpose of these projects: to proof that this increased level of simulation-complexity can yield the same results obtained in simpler simulations, while offering much more possibilities to investigate relative reproductive advantages of certain behavioural strategies. The current projects are intended as a proof-of-concept for doing research in evolutionary origins of behaviour with simulations in Framsticks. Only by doing many projects biologists can relate with, they can start to take these simulations seriously. This is important, for this kind of synthetic methodology to biological problems requires people from both sides to respect each other. Only then, we can start a fruitful dialogue and cooperation! Why do research in theoretical biology,
A: Theoretical biology models natural phenomena. Evolutionary robotics, on the other hand, is an engineering technique to automatically construct robot controllers. Virtual life research is not aimed at generating new models for biological phenomena (and is most probably not able to do so anyway). It is currently trying to gain recognition for the fact that these simulations can help us test and understand the reproductive advantages of certain behavioural patterns, in various environmental conditions. Thus, is can help testing hypothesis from theoretical biology. The primary goal is to understand the evolutionary origins of behaviour. This is a biological question, which we try to answer using engineering techniques from evolutionary robotics. And perhaps, results from these studies can help to improve the techniques used in evolutionary robotics. This, however, is considered a secondary target. Genetics not important for evolution ?!?
A: Virtual life research indeed makes the (huge) assumptions that behaviour is inheritable property. Moreover, we greatly simplify the evolutionary process by cancelling complicated genotype - phenotype mapping, such as developmental (epigenetic) processes. Of all the genes that newly borns get from their parents, only some are subject to change (mutation). Most importantly for current project are the genes that code for weights of connections in the neural network (the brain). Therefore, we can pretend that the only genes in the genotype are those that code for this property. In the near future, however, in the project of co-evolutionary complexicifation this changes. In this projects, the topology of the neural network can changes (grow, shrink, change structure) by structural mutations. To enable genetic operations (like crossover) to still result in offspring that inherits some of the behavioural (!) properties of its parents, new genotype encodings have to be specified. Although the genotypical encodings and mappings do not (yet) play a major role in Virtual life projects, it is of major importance to the Framsticks project itself. Different encodings (incl. one that incorporates developement) have been invented and implemented. For more information on this topic, refer to the publications of Maciej Komosinski, Szymon Ulatowski and others. Especially relevant publications are: - M. Komosinski, A. Rotaru-Varga, Comparison of Different Genotype Encodings for Simulated 3D Agents . In: Artificial Life Journal , 7 (4), 2001. MIT Press, 395-418. - M. Komosinski, M. Kubiak, Taxonomy in Alife. Measures of Similarity for Complex Artificial Organisms . In: Proceedings of 6th European Conference on Artificial Life (ECAL01) , September 10-14, 2001, Prague, Czech Republic, Springer-Verlag (LNAI 2159), 685-694. |