an experiment with evolution – developing an eye michael guzman 21/2/2007
TRANSCRIPT
An Experiment with An Experiment with Evolution Evolution – –
Developing an EyeDeveloping an Eye
Michael Guzman
21/2/2007
Outline Outline
• Introduction and Background
• Project details
• Selected results
• Summary
Introduction and BackgroundIntroduction and Background
• Genetic Algorithm
• Eye evolution – What we know by now
• Learn optics in 30 seconds
Genetic AlgorithmGenetic Algorithm
• Genetic algorithm is a probabilistic search algorithm.
• Iteratively transforms a population of individuals, each with an associated fitness value, into a new population of offspring objects.
• Darwinian principle of natural selection
• Applying operations which imitate nature’s genetic operations, such as crossover (sexual recombination) and mutation.
Genetic AlgorithmGenetic Algorithm
Eye evolution – What we knowEye evolution – What we know
Learn optics in 30 secondsLearn optics in 30 seconds
Learn optics in 30 secondsLearn optics in 30 seconds
Project detailsProject details
• Representation – The Genome
• Assumptions and Constants
• Fitness function
Representation – The GenomeRepresentation – The Genome
The purpose is to assume nothing
• The shape of the eye will we an ellipse – A-axis, B-axis
• The width of the opening to let light in
• Lens width
• Lens vertical location
• Lens focal length
Representation – The Representation – The GenomeGenome
A
B
Lens Y
Lens Width
Opening
Assumptions and ConstantsAssumptions and Constants
• Mirror symmetry• We start with very small B-axis almost a flat
patch with the widest opening possible• The starting focal length is very big – same
as starting with no lens.
• Mutation probability 5%• Mutation magnitude 5%• Crossover probability 90%
Fitness functionFitness function• We consider the following factors
1. The smearing of a point on the retina
2. Area πAB
3. Perimeter – Ramanujan approximation
4. Illumination power - (2×L-radius)^2/focal^2
5. Resolution – different points projected on different photoreceptors.
6. Opening size – how much light goes in
Fitness functionFitness function
• All factors normalized by their max-value
• The maximal values for the axes are 4 times bigger than the biggest eye existing today.
Selected resultsSelected results
• Why selected results?
Selected resultsSelected results – – eye #1eye #10
500
2000
5000
Selected resultsSelected results – – eye #2eye #20
1000
5000
Selected resultsSelected results – – eye #3eye #3
• Flat wide eye
• Minimal opening
• Lens adjacent to retina
• Very big focal length
What went wrong?
SummarySummary• Using an unconditioned (almost) model of the eye,
the results are nevertheless reasonable, and similar eyes can be found in nature.
• The project tries to simulate natural process from nature and therefore imposes some initial conditions on the individuals, a fact which prevents the genetic algorithm to show it full power.
• Some of the result are very improbable and they occur because of the method used to select the “parents” in each generation.
• It seems that otherwise than in size, no better eye than those we know from nature, has developed during the running of the algorithm.
Future work
• Finding a better general fitness function, giving more weight to : usage of the retina, ratio between axes etc…
• Trying special fitness function according to environmental conditions.
• Making a interactive web applet incorporating all of these .
ReferencesReferences• IBCV 2007 LectureNotes• Evolutionary Computation and Artificial Life - BGU course Le
cture• Wikipedia• A pessimistic estimate of the time required for an eye to
evolve Nilsson & Pelger.• Feynman lectures on physics Vol I ch.31• Field guide to Visual and Ophthalmic Optics bgu-lib QP.475.S385
• Mathematics handbook by Korn&Korn McGRAW-HILL
• Various physics and analytical geometry books (russian)