Tasks 1-2-3 : how can they be integrated together ?
Wenguo
Genome: represent an organismEvert's approach: evolution let's choose.
Abstract class Shape.
C++- cell.h
- have the same name
Communicate with people developing the controller first.
What kind of modules they need.
How the shape;
Once they have the shape.
From two genomes of shape, build the organism.
From 1st simulator, use Robot3D.
- Robot3D: Yao yes; Wenguo no;
Reached final simulation.
For us, we have to have sensors;
Robot3D;
If we have to use a single one;
- Wenguo's Stage.
Group 1.
Genotype to Phenotype
mapping one to many; many to one ?Not very important; but useful information.
Init. and Crossover
Wenguo: implemented: code available. simple examples.Yao-yao: involve the shape from scratch. Local topology.
The genome is a function, disseminated on every cell and deciding whom to recruit.
Describe coding.
summary: 1 explicit; 3 implicit representations.
Group 2. Jurgen.
- criteria to compare: agreed in skype meeting.
- I, T, H shapes
- base class which are evolved.
- share one symbol seed file (html); make comparison easy.
Graz: Online evolved AHHS. We have to simulate in 3D;
Can be used also for group 1.
Group 3: Internal reward.
- Evert was sick. will do it.
and then INRIA can return weights.
Group 4: Libor.
Simulation: 3 x slower than real-time, with 10 robots and sensors off.If you increase the number of cells; becomes 10 times slower.
Speed of simulation does not depend much on structure: only on number of cells.
1 hour: 10 minutes of real-time.
How many evaluations ?
Depends how fast an egg finds another cell.
Distinguish time in real-world / cycles in simulator
Vojta is preparing the manual, with Anne. Something early next week.
What about the installation ?
Having a place where to find tricks and recipes ?
- In e-mails to Libor;
- But hard to catch up if you enter the game now
- having a wiki ? On Symbrion, pages are outdated.
- use "longebran" (here I did not follow).
Agreed:
- Team 1 and 2, cooperate.
- Every leader prepares a report.
A meeting to reach a decision in December ?
Jurgen: hard to get the time from teaching.
After mid December: 19-23.
Guzs: I could attend iff in Amsterdam.
What about early January.
VU: closed in the week between Christmas and New Year.
Subtask2: Organism control parameters
On robot3D: should be running, dec. 15th.Evert: find on light simulator first.
Nicolas: until Nov. 30: light simulator, then Robot3D.
Jurgen: too many pbs to debug 30 nov; 15 dec.
Florian: But how can we compare if we don't have same simulator.
Jurgen: Postpone decision until 15 dec.
Yaoyao: need help about simulator
Vojta/Libor: need help from Lutz. One day meeting ? Skype call? Write documentation on simulator.
Lutz: you can send me questions.
Guzs: start soon with small tutorial. Then another one.
Skype tutorial on next week;
Subtask3: Evert & Michele & Christopher
Adapt an implementation of the QI.I plan to do that before the week-end.
Static version of QI.
Discussion after the skype.
Subtask 1: Morphogenesis
Wenguo: improved the representation: for next week, create mutation, crossover.send them to Christopher.
Yao: done some experiments in its simulator; working with Ronny to integrate his findings.
Michele: Before sunday evening, paragraph summarizing progress.
Subtask 4: Simulator
Distribute threads on computers ?Lutz: possible; not done yet.
Physics is the bottleneck.
Jurgen: two kernels; we don't use the threads. Is there anything we can read so that all kernels are used by default ?
Lutz: done in new code.
Nicolas: compiled on Ubuntu 11 ?
Ask Anne. support 10.10 and 11.04.
Florian: I run 11.04. happily.
What the first tutorial should say:
- how do we get; compile; run in the system; find good parameters.
- some examples which do not match.
Anne: packages are compiled every night.
Deadline for benchmark: Nov. 15th.
Scenario with snakes and simple creatures.
Guzs: but we have two more weeks for preliminary work. On Nov. 30th, all go to Robot3D.
Libor: ok for next week the benchmark; but the bugs, after.
Berend: we need a bug tracking system. Anne.
Would it be easier to have all the same version ? 11.04 if possible. Would simplify the debugging.
But do we have a machine where to run the compiler now ?
Agreed, bugs supported on 11.04 only.
Next
Each leader writes a paragraph on the progress before Sunday night.How to synchronize the expe in each subtask :
so that people do comparable experiments: private skype.On subtask 3 (internal reward): private skype right w
General question
Simulator: a paper of it ? Might be.How to debug it, yes; or the simulator.
Already a paper, Symbricator 3D.
The software might be a legacy to the community. Unique properties.
Group Morphogenesis
Task
Decide for a representation of shapesPartners
- Wenguo (Bristol)
- Explicit representation, grammar-like
- Tested in simulation: organism construction completed
- Issue: evolvability
- Ronny (Graz)
- Virtual embryogeny, AHHS
- Each module of the organism exchanges signals (hormone levels ?) and surface module can recruit additional free cells
- Tested in simulation () and on robots (how many)
- Issue: size of the representation
- Michele (Paris)
- Virtual embryogeny, cellular automaton
- Each module of the organism emits signals and surface modules can recruit additional free cells
- Tested in simulation () PhD Alexandre Devert
- Issue: convergence toward controllable-size organisms
- Christopher (Tubingen)
- Helps Wenguo to devise structured variation operators (mutation, crossover)
- Yao-yao (Gent)
- Suggestions to Ronny.
Criteria of evolvability
A viable shape
is an organism which includes 4-10 modules and does not recruit anyone else.tbcompleted.
Random initialization
Using random descriptions in the representation space- Number of trials needed to yield XX viable and diverse shapes
- Time needed
Controllability
Feasibility of describing a prescribed shape in the representation spaceEfficiency of operators
Using (crossover, mutation) on viable parent shapes,- percentage of trials needed to yield XX viable offspring
- Time needed
Ratio: Time to reach a viable shape from scratch /
Time to reach a viable shape from two viable parent shapes.
Experimental setting
Overview
- Initialization (from fixed shapes or from random)
- Loop:
- Variation operators
- Fitness = number of new viable shapes offspring
- End loop
Later
- Fitness will be enriched with other criteria (walking).