Report from the decision meeting (11) Jan

Answer from Graz (Skype meeting, Jan. 13th and mail 15th of January)

  • Size of the genome (Virtual Embryogenesis): between 50 and 500 depending on the shape complexity
  • Viability: 5 genomes viables out of 1,000 random genomes.
  • Mapping is deterministic (except for the fact that there might be no free robots in the environment)
  • Controllability not tested but some examples of shape attached.
  • Max size: goes to 10 after 150 generations on 5 runs (Runs attached). Question, how many individuals per generation ? Let's assume 30.
  • Number of different shapes: not yet.

Answer from Ghent (Skype meeting, Jan. 13th)

  • Size for a shape of size n: 3n
  • A threshold is used; default value considered in the following is .3 for each gene.
  • Out of 1,000 random genomes (with .3 threshold), 80% are viable; 40% are unique, viable shapes
  • Mapping stochastic; ratio phenotype/genotype depends on a threshold; for threshold .3, figure will arrive.
  • Controllability: examples of shape will arrive shortly.
  • Initialization: choose the size; then you have a string; compare each value to the threshold; this makes it an explicit representation.

Next step:

  • Complete the figures for the report
  • Clarify relationship between GRN and AGE.

Summary of the Jan. 11th

Regarding our Morphogenesis task, it was decided that *for the demo* we would consider one explicit representation (Wenguo's is chosen as the most advanced one) and one implicit representation. All representations will be considered
in view of the report and the paper.

There are 3 implicit representations at the moment: Graz's (Virtual Embryogeny), Paris' (Cellular Automaton) and Ghent's (Yao's GRN).

The decision was not taken on the spot out of lack of information: some measurements were missing on GRN and VE.

These measurements are the following (nothing fancy; the usual tables and figures found in every experimental section of every paper in evolutionary computation):

  • what is the size (number of symbols, integers, real-values) of the genotype needed to encode a shape with n robots ? If the genotype to phenotype mapping involves the environment (e.g. the sensors of the robots), consider an average environment/empty environment in Robot3D for the sake of reproducibility.
  • what is the proportion of viable (i.e. less than 10 robots, not 4 robots linked in a square) built from 1,000 genomes drawn uniformly in the search space ?
  • is the mapping (genotype to phenotype) deterministic or stochastic ? In the latter case, how many viable shapes (phenotypes) are built from one genotype on average ?
  • what is the controllability of the representation, i.e.: consider a stair or a cross shape, is there a genotype coding for these shapes ?
  • what is the evolvability of the representation, i.e., take a random population, use what you like as fitness (my recommandation, the size of the organism) and draw
  • ** the max. size of the organism built vs the number of fitness evaluations (average out of 10 runs);

** the number of different shapes (two shapes are equal if they can exactly be super-imposed, egg on egg) vs the number of fitness evaluations (average out of 10 runs);
** the number of different shapes (two shapes are equal if they can exactly be super-imposed while keeping the egg) vs the number of fitness evaluations corresponding to viable shapes (average out of 10 runs);

Unless otherwise stated, the content of this page is licensed under Creative Commons Attribution-NonCommercial-ShareAlike 3.0 License