April 29, 2017
 

Autonomic Systems

Evolving virtual embryogenesis to structure complex controllers

Ronald Thenius, Michael Bodi, Thomas Schmickl, and Karl Crailsheim
Virtual evolution helps increase the efficiency of classical controller paradigms by implementing substructures with different tasks and capabilities.

To generate complex behaviours in autonomous technical units such as autonomous robots, controllers must be developed that can respond to unknown situations (e.g., during rescue missions in collapsed buildings or while exploring other planets). Development of such autonomous systems is a topic of intense current research.1,2

With increasing complexity of the tasks, ‘classical’ controllers, such as artificial neural networks, soon become inefficient. This happens because of the increasing complexity of the learning or adaptation processes. To enhance the adaptivity of such classical controller paradigms, one can implement substructures aimed at performing different tasks and with varying capabilities (i.e., heterogeneous controller structures). Structuring of the controller can be predefined manually3,4 or shaped through artificial evolution. To automatically develop such structures, we need to use a process that is easy to evolve and rich in different solutions.

One inspiration for such developments comes from biological embryogenesis processes. Throughout natural evolution, embryogenesis has proved to be an ideal tool for shaping the bodies (including the control structures) of multicellular lifeforms. The underlying mechanisms are known as EvoDevo.5 They are perfectly suited to operate as a framework for evolutionary processes. To reach this goal in an artificial-life context, we defined the concept of virtual embryogenesis (VE).6 By employing VE in combination with artificial evolution (AE), we developed a new method for controller structuring (see Figure 1).


Embryo adaptation throughout the evolutionary process. (a) Desired shape of the final embryo (part of the fitness function of the artificial evolutionary process). The same desired shape can, if required, be used to develop multilayer artificial neural networks (ANNs). (b)–(f) Change of body shape as a function of evolutionary stage. Dots indicate single cells within the embryo.

Growth of biological embryos is controlled by morphogens: see Figure 2(a). Cells respond to different concentrations of morphogens, e.g., by duplication or emission of morphogens. Cell response is encoded in the genome. By modelling these processes—see Figure 2(b) and (c)—we can develop controller structures—see Figure 2(d)—and test them against predefined benchmarks.6


Biological and virtual embryos. (a) Different morphogen concentrations in a biological embryo.7(b) Morphogen concentrations in a hand-coded virtual embryo. Dots and colours indicate single cells and morphogen gradients, respectively. (c) Virtual embryo shaped by artificial evolution. Colours indicate a morphogen gradient that is responsible for lateral growth. (d) Structured ANN developed inside a virtual embryo. Dark/light lines indicate short/long-distance connections (local subnetworks/global network).

We next plan to investigate possibilities to increase the potential of VE by optimizing the interplay between AE and VE. We expect to increase the quality of the evolved structures by implementing gene-duplication and transposition processes. In addition, we will adapt VE to controller types used in multimodular robotics4 and investigate its benefits for these controller paradigms.


Authors

Ronald Thenius
University of Graz

Ronald Thenius has been a member of the Artificial Life Laboratory since 2008. He is currently working on development of new control paradigms for autonomous robotic systems within the European Union projects REPLICATOR and SYMBRION.

Michael Bodi
University of Graz

Thomas Schmickl
University of Graz

Karl Crailsheim
University of Graz


References
  1. http://www.symbrion.eu SYMBRION project website. Accessed 14 September 2010.

  2. http://www.replicators.eu REPLICATOR project website. Accessed 14 September 2010.

  3. S. Nolfi and D. Parisi, Auto-teaching: networks that develop their own teaching input, Proc. 2nd Eur. Conf. Artif. Life, 1993.

  4. M. Neal and Jon Timmis, Timidity: a useful emotional mechanism for robot control?, Informitica 27 (2), pp. 197-204, 2003.

  5. G. B. Müller, Evo-Devo: extending the evolutionary synthesis, Nat. Rev. Genet. 8, pp. 943-949, 2007.

  6. R. Thenius, T. Schmickl and K. Crailsheim, Novel concept of modelling embryology for structuring an artificial neural network, Proc. 6th Vienna Int'l Conf. Math. Model. (MathMod), 2009.

  7. J. Jaeger, S. Surkova, M. Blagov, H. Janssens, D. Kosman, K. N. Kozlov, Manu, E. Myasnikova, C. E. Vanario-Alonso, M. Samsonova, D. H. Sharp and J. Reinitz, Dynamic control of positional information in the early Drosophila embryo, Nature 430, pp. 368-371, 2004.


 
DOI:  10.2417/2201009.003291