Neural Optimization of Evolutionary Algorithm Strategy Parameters

Neural Optimization of
Evolutionary Algorithm
Strategy Parameters
Hiral Patel
Outline
Why optimize parameters of an EA?
 Why use neural networks?
 What has been done so far in this
field?
 Experimental Model
 Preliminary Results and Conclusion
 Questions

Why optimize parameters of
an EA?
 Faster
convergence
 Better overall results
 Avoid premature convergence
Why use neural networks?
 Ability
to learn
 Adaptability
 Pattern recognition
 Faster then using another EA
What has been done so far in
this field?
 Machine
Learning primarily used to
optimize ES and EP
 Optimized mutation operators
 Little has been done to optimize GA
parameters
Experimental Model Outline
 Neural
Network Basics
 Hebbian Learning
 Parameters of the Genetic
Algorithm to be optimized
 Neural Network Inputs
Neural Network Basics
bq(k) bias
Vector input
signal
x(k)Rn1
wq1(k)
wq2(k)
Sigmoid
activation
function
Neuron
response
(output)
yq(k)
f(•)
vq(k)
wqn(k)
Synaptic
weights
x(k)Rn1
Deviation of
activation
function
g(•)=f’(•)
dq(k)
Desired
neuron
response
e (k )
Weight update algorithm
Adapted from: Ham, M. H., Kostanic, I Principles of Neurocomputing for Science and Engineering, McGraw-Hilll, NY, 2001
Hebbian Learning
 Unsupervised
learning
 Time-dependent
 Learning signal and Forgetting
factor
x0
Hebb Learning for
single neuron
w0
x1
xn
w1
v
f(v)
y
wn
Standard Hebbian
learning rule
{,}
ly
d (v)
 f (v )
dv
Adapted from: Ham, M. H., Kostanic, I Principles of Neurocomputing for Science and Engineering, McGraw-Hilll, NY, 2001
Parameters of the Genetic
Algorithm to be optimized
Crossover Probability
 Crossover Cell Divider
 Cell Crossover Probability
 Mutation Probability
 Mutation Cell Divider
 Cell Mutation Probability
 Bit Mutation Probability

Neural Network Inputs
Current Parameter Values
 Variance
 Mean
 Max fitness
 Average bit changes for crossover
 Constant parameters of the GA

Preliminary Results
Tests run with Knapsack problem with
dataset 3, pop. size 800, rep. size 1600
 Learning Signal and Forgetting factor
are not yet optimal enough to suggest
better performance with NN

Output for 1600 generations
500
400
Fitness
300
Mean
200
Variance
CCD
100
MCD
0
0
-100
500
1000
1500
2000
Probabilities for 1600
generations
1.2
1
CP
0.8
CCP
0.6
MP
CMP
0.4
BMP
0.2
0
0
500
1000
1500
2000
Conclusion

It may be possible to get better
performance out of a Neural Optimized
EA as long as the (unsupervised) Neural
Network is able to adapt to the changes
quickly and to recognize local minima.
Possible Future Work

ES to optimize parameters, use a SOM to
do feature extraction of the optimized
parameter values, use the SOM output as
codebook vectors for LVQ network and
then classify the output of the original ES,
use the classifications to perform
supervised training of LevenbergMarquardt Backpropagation network to
form rule set.
Question ?