Download slides - Stefano Nichele

1
Evolutionary Growth of Genomes
for the Development and Replication
of Multi-Cellular Organisms with Indirect Encodings
Stefano Nichele and Gunnar Tufte
ICES 2014 - Orlando (USA)
December 9-12, 2014
Stefano Nichele, 2014
2
Genomes of biological organisms are not fixed in size
They evolved and diverged into different species acquiring
new genes and thus having different lengths
• LUA (Last Universal Ancestor) ~ 3.5 / 3.8 billion years ago
• Gene duplication: redundant gene with less selection pressure
• Larger genome: genetic novelty, potential for innovation
• Complexification: incremental elaboration
• ~ 38% Homo Sapiens genome due to gene duplication
3
Artificial EvoDevo systems often have static size genomes
System designer choice:
• Trial & error
• Estimation / heuristics
Fixed maximum complexity (vs. open-ended in nature)
4
Goal
• Evolutionary growth of genomes (with indirect encodings)
– Initialize genomes with a single gene (low dimensionality)
– Allow gene duplications (add new degrees of freedom)
• Evo-Devo System based on Cellular Automata
– Abstract model of development
– Morphogenesis
– Replication
• Evolve compact and effective genomes
– Compare genome size and success rate
– Fully specified genome (complete CA transition tables) vs growing genome
5
EvoDevo Mappings
• Direct
• Redundant (neutrality)
• Indirect (generative/developmental)
– Full specification of representation
– von Neumann replicator: 295
• Fixed length (subset)
• Variable length
• Complexification
– NEAT (Stanley & Miikkulainen): good for evolving modular structures with
direct encodings
6
CA as EvoDevo systems
•
CA can be considered as a developmental system, in which an organism can
develop (e.g. grow) from a zygote to a multi-cellular organism (phenotype)
according to specific local rules, represented by a genome (genotype).
•
The behavior of the CA is represented by the emergent phenotype, which is
subject to shape and size modification, along the developmental process.
7
Traditional CA dev. model
Example CA with 4 cell states and 5 neighbors:
Search space = 4^4^5 = 41024 = ~ 3.23 x 10616
8
Previous work
• Evolutionary growth of genomes
– CA transition tables
– abstract measures of complexity: trajectory / attractor length
• Scalability
– Search space, number of cell states, geomerty size (phenotypic resources)
Current:
• Different target: phenotypic structures of different complexity
• Different mapping: IBD (Instruction-Based Development)
– Not bounded (evolve from one instruction to program)
9
Evolutionary Growth
• A genetic algorithm (details in paper) with 4 regulation
mechanisms to control gene duplication:
–
–
–
–
Upper bound, total number of genes
Duplication rate
Optimization time (before new duplication can occur)
Elitism
• Weighted fitness:
– 80% actual fitness
– 20% innovation parameter
•
•
Rewards larger genomes
New genes most likely fitness-neutral or negative
10
CA – IBD
(Bidlo and Skarvada 2008, Bidlo and Vasicek 2012)
U
L
C R
D
gene
Inst.
Code
Op1
Op2
11
Benchmark structures
12
Development problem
6 1 3
3 3 1
Inst.
Code
8 3 4
Op1
11 4 3
6 3 0
Op2
2 4 3
1 1 4
2 1 4
6 4 1
15 4 3
5 2 0
1 2 4
13 4 0
Operands: U = 0, R = 1, D = 2, L = 3, C = 4.
0 2 0
13
Development - results
Replication problem
time
14
15
Replication - results
16
Replication of ”French” Flag
17
Success rate
Development (avg. 100 runs)
Replication (avg. 100 runs)
Table-based Evolution
Success
Rate %
Table-based Evolution
Genotype Size (# genes)
Max
Avg
Min
StDev
Generations
Avg. StDev.
Success
Rate %
Genotype Size (# genes)
Max
Avg
Min
StDev
Generations
Avg. StDev.
A
58
32
32
32
0
1336
2294
A
85
32
32
32
0
775
1393
B
69
32
32
32
0
2254
2501
C
8
1024
1024
1024
0
4331
3576
C
19
1024
1024
1024
0
5002
3157
D
1
32
32
32
0
8259
0
D
23
32
32
32
0
2668
2942
E
0
1024
1024
1024
0
-
-
Instruction-based Growing Evolution
Success
Rate %
Instruction-based Growing Evolution
Genotype Size (# genes)
Max
Avg
Min
StDev
Generations
Avg. StDev.
Success
Rate %
Max
Genotype Size (# genes)
Avg
Min
StDev
Generations
Avg. StDev.
A
98
31
14.34
5
8.4318
1257
1152
A
100
7
2.93
2
1.1742
39.7
19.6
B
98
31
15.28
5
7.0973
3956
1690
C
100
6
2.84
2
1.1166
39.6
22.3
C
46
46
19.65
6
9.2236
6424
1922
D
100
8
3.06
2
1.2128
41.8
20.5
D
100
13
5.25
4
1.4097
285
108
E
100
5
1.38
1
0.8012
9.4
10.7
18
Conclusion
• Evolutionary growth of genome
• Initialize with single gene, allow duplication and
speciation (regulation mechanisms)
• Traditional CA mapping vs instruction based
development (unbounded)
• Development and replication problems
• Compact and effective genotype solutions (not
designed a priori)
• Better success rate
19
Future work
• Investigate a major challenge in EvoDevo:
– Development of complex morphologies and structures
(potentially at levels of complexity found in nature)
• True complexification
– Allow growth of available cell states (unbounded state space)
• Optimization of instructions and instruction set
• More benchmarks and tasks (example: circuit design)
• Introduce self-modifying instructions
– Allow diversification of programs
20
Stefano Nichele
www.nichele.eu