An Examination of Synchronisation in Artificial Gene Regulatory Networks
Jonathan Byrne, Miguel Nicolau, Anthony Brabazon and Michael O'Neill
Natural Computing Research and Applications Group
University College Dublin
Ireland
Overview
The gene regulatory network (GRN) model
Synchronisation step
Pole balancing experiment
Offset experiments
Conclusions
Gene Regulatory Network
Gene expression generates proteins
Proteins enhance or inhibit gene expression
DNA segments indirectly interact through protein production
Banzhaf GRN model
Evolutionary: Evolved bitstring model
Developmental: Fully connected Network
Evo Representation
Defines network structure
Promoter defines number of nodes
Gene sequence defines the weights
Evo Representation
Devo Representation
Fully connected network topology
Node interactions defined by weights
Node concentrations change over time
Devo Representation
GRN Input and Output
"Extra" genes used for input
"P" genes used for output
Extra concentration set by input value
Extra limited to a maximum of 0.4
Extra and TF genes normalised together
P genes normalised separately
GRN Input and Output
The Synchronisation Step
Change in input effects dynamics of the network
Stabilisation required to react to input
How does the size of the sync step effect the GRN?
Is there an automatic approach for choosing the step size?
Pole Balancing Experiment
Pole Balancing Experiment
Dynamic experiment benchmark
120,000 time steps
Cart re-initialised for every generation
4 Extra genes and 2 P genes
Pole Balancing Experiment
X ∈ [−2.4, 2.4] m is the position of the cart from the centre
θ ∈ [−12, 12] ◦ is the angle of the pole with the vertical
x ∈ [−1, 1] m/s is the velocity of the cart on the track
Θ ∈ [−1.5, 1.5] ◦ /s is the angular velocity of the pole
Pole Balancing Results
Offset Experiments
Same input with different offsets for target output
Can it match the target?
Does increasing the size of the sync step effect performance?
Does a stabilised network retain information about previous states?
Sine Experiment
Offset 10
Offset 20
Sine Results: Offset 0
Sync step 10 offset 0
Sine Results: Offset 10
Sine Results: Offset 20
Random Walk Experiment
Random Walk Results: Offset 0
Random Walk Results: Offset 10
Random Walk Results: Offset 20
Conclusions
Synchronisation can improve performance
Improvement increased logarithmically
No improvement with greater offsets
Variable synchronisation performed worse
Questions?