evolutionary origins of modularity02317/slides/lec_18.pdf · evolvability •evolvability...
TRANSCRIPT
![Page 1: Evolutionary origins of modularity02317/slides/lec_18.pdf · Evolvability •Evolvability –capacity to rapidly adapt to novel environments •Two organisms with the same phenotype](https://reader034.vdocuments.site/reader034/viewer/2022042921/5f68cbbb1e4a6703dc101398/html5/thumbnails/1.jpg)
Evolutionary origins of modularity
Jeff Clune, Jean-Baptiste Mouret and Hod LipsonProceedings of the Royal Society B 2013
Presented by
Raghav Partha
![Page 2: Evolutionary origins of modularity02317/slides/lec_18.pdf · Evolvability •Evolvability –capacity to rapidly adapt to novel environments •Two organisms with the same phenotype](https://reader034.vdocuments.site/reader034/viewer/2022042921/5f68cbbb1e4a6703dc101398/html5/thumbnails/2.jpg)
Evolvability
• Evolvability – capacity to rapidly adapt to novel environments
• Two organisms with the same phenotype and fitness in a current environment may differ in their evolvability
How does evolvability arise?
![Page 3: Evolutionary origins of modularity02317/slides/lec_18.pdf · Evolvability •Evolvability –capacity to rapidly adapt to novel environments •Two organisms with the same phenotype](https://reader034.vdocuments.site/reader034/viewer/2022042921/5f68cbbb1e4a6703dc101398/html5/thumbnails/3.jpg)
Modularity contributes to evolvability
• Engineering – easier to design/rewire modules
• Biological entities are modular• Brain
• Metabolic Pathways
• Protein interactions
Why does modularity evolve?
![Page 4: Evolutionary origins of modularity02317/slides/lec_18.pdf · Evolvability •Evolvability –capacity to rapidly adapt to novel environments •Two organisms with the same phenotype](https://reader034.vdocuments.site/reader034/viewer/2022042921/5f68cbbb1e4a6703dc101398/html5/thumbnails/4.jpg)
Why does modularity evolve?
• (Indirect) selection for evolvability (modularity)
• Leading Hypotheses• Rapidly changing environments with common subproblems
![Page 5: Evolutionary origins of modularity02317/slides/lec_18.pdf · Evolvability •Evolvability –capacity to rapidly adapt to novel environments •Two organisms with the same phenotype](https://reader034.vdocuments.site/reader034/viewer/2022042921/5f68cbbb1e4a6703dc101398/html5/thumbnails/5.jpg)
Why does modularity evolve?
• (Indirect) selection for evolvability (modularity)
• Leading Hypotheses
• Rapidly changing environments with common subproblems
• Bacteria occupying diverse environments have more modular metabolic networks
Is there a simpler/testable hypothesis?
![Page 6: Evolutionary origins of modularity02317/slides/lec_18.pdf · Evolvability •Evolvability –capacity to rapidly adapt to novel environments •Two organisms with the same phenotype](https://reader034.vdocuments.site/reader034/viewer/2022042921/5f68cbbb1e4a6703dc101398/html5/thumbnails/6.jpg)
Selection to reduce connection costs leads to modularity?• Connection costs in networks
• Neural networks – metabolic / energy
• Signaling pathways – delay in output of a critical response
• Gene regulation – Limit on DNA binding sites
• Evidence for cost selection• Summed length of wiring diagram minimized in animal brains
![Page 7: Evolutionary origins of modularity02317/slides/lec_18.pdf · Evolvability •Evolvability –capacity to rapidly adapt to novel environments •Two organisms with the same phenotype](https://reader034.vdocuments.site/reader034/viewer/2022042921/5f68cbbb1e4a6703dc101398/html5/thumbnails/7.jpg)
Computational evolution of modularity
![Page 8: Evolutionary origins of modularity02317/slides/lec_18.pdf · Evolvability •Evolvability –capacity to rapidly adapt to novel environments •Two organisms with the same phenotype](https://reader034.vdocuments.site/reader034/viewer/2022042921/5f68cbbb1e4a6703dc101398/html5/thumbnails/8.jpg)
Computational Evolution of Modularity
• Each individual is a network that takes stimuli and returns an outputModular goal
Fitness of a network = Fraction of input stimuli the network gets correctModularity:
![Page 9: Evolutionary origins of modularity02317/slides/lec_18.pdf · Evolvability •Evolvability –capacity to rapidly adapt to novel environments •Two organisms with the same phenotype](https://reader034.vdocuments.site/reader034/viewer/2022042921/5f68cbbb1e4a6703dc101398/html5/thumbnails/9.jpg)
Two simulation experiments
• Connection Costs: Distance between the nodes
• Compare the modularity of optimal networks evolved in the two simulations
PAMaximizing Performance
alone
P&CCMaximizing Performance
andMinimizing Connection Costs
![Page 10: Evolutionary origins of modularity02317/slides/lec_18.pdf · Evolvability •Evolvability –capacity to rapidly adapt to novel environments •Two organisms with the same phenotype](https://reader034.vdocuments.site/reader034/viewer/2022042921/5f68cbbb1e4a6703dc101398/html5/thumbnails/10.jpg)
PA evolution
![Page 11: Evolutionary origins of modularity02317/slides/lec_18.pdf · Evolvability •Evolvability –capacity to rapidly adapt to novel environments •Two organisms with the same phenotype](https://reader034.vdocuments.site/reader034/viewer/2022042921/5f68cbbb1e4a6703dc101398/html5/thumbnails/11.jpg)
P- CC: Non-dominated Sorting Genetic Algorithm
• Stochastic Pareto Dominance – use CC only 25% of the time1. To select parent for mutation, leading to offspring
2. To select N fittest individuals for the next generation
![Page 12: Evolutionary origins of modularity02317/slides/lec_18.pdf · Evolvability •Evolvability –capacity to rapidly adapt to novel environments •Two organisms with the same phenotype](https://reader034.vdocuments.site/reader034/viewer/2022042921/5f68cbbb1e4a6703dc101398/html5/thumbnails/12.jpg)
Results
• 25,000 iterations/generations of evolution with fitness calculated using the two different conditions
• Mutations – randomly adding or removing edges from the network; modifying weights on edges, bias on the nodes of neural network
PAMaximizing Performance
alone
P&CCMaximizing Performance
andMinimizing Connection Costs
50 trials each
![Page 13: Evolutionary origins of modularity02317/slides/lec_18.pdf · Evolvability •Evolvability –capacity to rapidly adapt to novel environments •Two organisms with the same phenotype](https://reader034.vdocuments.site/reader034/viewer/2022042921/5f68cbbb1e4a6703dc101398/html5/thumbnails/13.jpg)
![Page 14: Evolutionary origins of modularity02317/slides/lec_18.pdf · Evolvability •Evolvability –capacity to rapidly adapt to novel environments •Two organisms with the same phenotype](https://reader034.vdocuments.site/reader034/viewer/2022042921/5f68cbbb1e4a6703dc101398/html5/thumbnails/14.jpg)
P-CC evolution produces more modular networks
![Page 15: Evolutionary origins of modularity02317/slides/lec_18.pdf · Evolvability •Evolvability –capacity to rapidly adapt to novel environments •Two organisms with the same phenotype](https://reader034.vdocuments.site/reader034/viewer/2022042921/5f68cbbb1e4a6703dc101398/html5/thumbnails/15.jpg)
P-CC evolution produces more modular networks
• Best performing network becomes more and more Modular
![Page 16: Evolutionary origins of modularity02317/slides/lec_18.pdf · Evolvability •Evolvability –capacity to rapidly adapt to novel environments •Two organisms with the same phenotype](https://reader034.vdocuments.site/reader034/viewer/2022042921/5f68cbbb1e4a6703dc101398/html5/thumbnails/16.jpg)
Is the modularity functional?
• Are the Left and Right subnetwork inputs in separate modules?• 56% of networks evolved under P&CC
• 0 networks evolved under PA
![Page 17: Evolutionary origins of modularity02317/slides/lec_18.pdf · Evolvability •Evolvability –capacity to rapidly adapt to novel environments •Two organisms with the same phenotype](https://reader034.vdocuments.site/reader034/viewer/2022042921/5f68cbbb1e4a6703dc101398/html5/thumbnails/17.jpg)
P-CC produces better networks
Generation
![Page 18: Evolutionary origins of modularity02317/slides/lec_18.pdf · Evolvability •Evolvability –capacity to rapidly adapt to novel environments •Two organisms with the same phenotype](https://reader034.vdocuments.site/reader034/viewer/2022042921/5f68cbbb1e4a6703dc101398/html5/thumbnails/18.jpg)
Why do P-CC networks achieve higher performance than PA networks?
• P&CC has additional constraint (minimizing costs) but still does better in terms of performance than PA
• Mutational effects are smaller, restricted to subcomponents
• Fewer connections – fewer parameters being optimized• Faster optimization
![Page 19: Evolutionary origins of modularity02317/slides/lec_18.pdf · Evolvability •Evolvability –capacity to rapidly adapt to novel environments •Two organisms with the same phenotype](https://reader034.vdocuments.site/reader034/viewer/2022042921/5f68cbbb1e4a6703dc101398/html5/thumbnails/19.jpg)
Why do P-CC networks are more modular and high-performing?• Inverse correlation between
Cost and Modularity for high
performing networks
• Existence of high performing
networks with low modularity
![Page 20: Evolutionary origins of modularity02317/slides/lec_18.pdf · Evolvability •Evolvability –capacity to rapidly adapt to novel environments •Two organisms with the same phenotype](https://reader034.vdocuments.site/reader034/viewer/2022042921/5f68cbbb1e4a6703dc101398/html5/thumbnails/20.jpg)
P-CC moves populations towards low cost high modularity solutions
PA populations
P-CC populations
![Page 21: Evolutionary origins of modularity02317/slides/lec_18.pdf · Evolvability •Evolvability –capacity to rapidly adapt to novel environments •Two organisms with the same phenotype](https://reader034.vdocuments.site/reader034/viewer/2022042921/5f68cbbb1e4a6703dc101398/html5/thumbnails/21.jpg)
Evolvability
• How evolvable are the networks that were obtained from the previous two simulation experiments?
• Which is suited better to adapt to a slightly different environment?
• L-OR-R vs L-AND-R
![Page 22: Evolutionary origins of modularity02317/slides/lec_18.pdf · Evolvability •Evolvability –capacity to rapidly adapt to novel environments •Two organisms with the same phenotype](https://reader034.vdocuments.site/reader034/viewer/2022042921/5f68cbbb1e4a6703dc101398/html5/thumbnails/22.jpg)
Modular networks are more evolvable
![Page 23: Evolutionary origins of modularity02317/slides/lec_18.pdf · Evolvability •Evolvability –capacity to rapidly adapt to novel environments •Two organisms with the same phenotype](https://reader034.vdocuments.site/reader034/viewer/2022042921/5f68cbbb1e4a6703dc101398/html5/thumbnails/23.jpg)
Modular networks are more evolvable
![Page 24: Evolutionary origins of modularity02317/slides/lec_18.pdf · Evolvability •Evolvability –capacity to rapidly adapt to novel environments •Two organisms with the same phenotype](https://reader034.vdocuments.site/reader034/viewer/2022042921/5f68cbbb1e4a6703dc101398/html5/thumbnails/24.jpg)
Computational Evolution of Modularity
![Page 25: Evolutionary origins of modularity02317/slides/lec_18.pdf · Evolvability •Evolvability –capacity to rapidly adapt to novel environments •Two organisms with the same phenotype](https://reader034.vdocuments.site/reader034/viewer/2022042921/5f68cbbb1e4a6703dc101398/html5/thumbnails/25.jpg)
Selection on connection costs leads to modularity
![Page 26: Evolutionary origins of modularity02317/slides/lec_18.pdf · Evolvability •Evolvability –capacity to rapidly adapt to novel environments •Two organisms with the same phenotype](https://reader034.vdocuments.site/reader034/viewer/2022042921/5f68cbbb1e4a6703dc101398/html5/thumbnails/26.jpg)
Summary
• Selection to reduce connection costs causes modularity, even in unchanging environments
• Got a lot of media attention for discovering ‘holy grail’ of evolving modular networks
• Significant advance in evolving modular design• Faster adaptation/evolvability