Evolving Modular Genetic Regulatory Networks

1
Evolving Modular Genetic Regulatory Networks

• By Josh Bongard, 2002
• Presented by Matt Luciw for
• CSE 848

2

• Ontogeny The development of an organism, from a
  single cell to mature form.
• This process is driven by the organisms
  underlying Genetic Regulatory Network (GRN).
• The GRN is copied into each cell as each is
  created.
• It determines the behavior (type) of each cell.

3

• Think of a GRN as a complex, directed network.
• Genes as nodes.
• Some are structural and directly contribute to
  cell behavior, when they are expressed.
• Others are regulatory and enhance or inhibit the
  expression of other genes.
• This paper presents a method for evolving GRNs
  called Artificial Ontogeny.

4
Why?

• Study the evolution of complexity from the
  bottom-up.
• Hypothesis A complex phenotype must be evolved
  hierarchically. Higher level genes will use
  lower level structures as building blocks.
• Example The Hox genes.
• Also A complex phenotype must be modular.
• Optimize a particular piece without interfering
  with other functionally-unrelated pieces.

5
Artificial Ontogeny

• 1) A 300 step growth phase, where the GRN
  develops the body (set of connected cylinders)
  and neural network controller of each agent,
  starting with a single cylinder.
• 2) A 500 step evaluation phase.
• Goal forward locomotion on an infinite
  horizontal plane.

6
AO Overview

• Physical structure
• One or more connected cylindrical units.
• Six diffusion sites.
• Neural network controller
• Sensors Touch, Proprioceptive (angle)
• Actuators Joint angle control (creates torque)
• Neurons Weighted sum, transfer function
• CPGs Output a sine wave.
• Bias Output a constant value.

7

• Note
• Six diffusion sites
• An initial motor neuron.
• Posterior and anterior axes.
• The genes from the initial genome.

8
The Genome

• A parser searches for promotor sites in the
  variable length genome to find genes.
• Each genome was initialized as 200 floating
  points between 0 and 1, and contained, on average
  10 promotor sites (10 genes).
• The values after the promotor site determine what
  transcription factor (TF) that gene produces and
  what (and how much) TF regulates its expression.

9

• Note
• Structural genes.
• Regulatory genes.
• Concentrations of transcription factors.

10
Examples of Expressed Gene Fn

• (this is from a different paper by the same
  author)

11

• Growth

12

• Ready for evaluation.

It runs in the simulated environment for 500 time
steps, then
13
Fitness

• Agents were biased (shaped) to have more units
  with actuated joints, and at least a
  minimally-connected neural network.

14
GA Summary

• 60 runs over 300 generations.
• Strong elitism 150 best retained each
  generation.
• Unequal crossover four variable crossover
  points.
• Parents chosen by tournament w/ size 3.
• For all new genomes
• 10 chance of substring deletion.
• 10 chance of substring swapping.
• On average, a single value is mutated.

15
Results

• Most agents that displayed locomotion were not
  that complex.

16
(No Transcript)
17
(No Transcript)
18
(No Transcript)
19
(No Transcript)
20

• (c) Loss of function
• (d) Gain of function

21

• (e) Differences in gene expression between LoF
  and original in the growth phase.
• (f) same for GoF.
• Dark grey structural neurological genes.
• Grey structural morphological genes.

22

• Quantitated neurological and morphological effect
  of these genes.
• Found first time of appearance of these genes for
  some complex agents.
• No initial morphological effect.
• Did these agents evolve to become complex because
  of this modularity?

23
Conclusions

• AO can develop locomoting agents with a high part
  count.
• These populations became successful due to early
  evolution of modular GRNs.
• (Correlation or causation?)
• Speculation do these genes function as master
  control genes, similar to the Hox genes?
• Evolution of GRNs by selection pressure.