Optimizing%20Models%20Using%20Continuous%20Ant%20Algorithms

1
Optimizing Models Using Continuous Ant Algorithms

• Oleg Kovárík
• kovaro1_at_fel.cvut.cz
• Pavel Kordík
• kordik_at_fel.cvut.cz
• http//cig.felk.cvut.cz
• Computational Intelligence Group
• Department of Computer Science and Engineering
• Faculty of Electrical Engineering
• Czech Technical University in Prague

2
Content

• GAME models
• Continuous Ant Algorithms
• Spiral data problem
• Interesting results

3
GAME
4
Types of units
5
Optimization methods

• QuasiNewton
• SADE
• PSO
• Hybrid GA-PSO
• Differential Evolution
• (Stochastic) Orthogonal Search
• Conjugate Gradient
• Ant algorithms AACA, ACO, CACO, DACO
• Random
• Powell
• ...

6
GAME
Select transfer function, parameters, optimizatio
n method for each unit
Genetic Algorithm
7
GAME
Genetic Algorithm
8
GAME
9
Ant Colony Optimization (ACO)?
Which next? Nearest? Used in best solutions?
Pheromone map
10
Direct Application of ACO

• DACO (Min Kong, Peng Tian 2006)?
• n variables xi with normal distribution
• N (µi , si ), i ? 1, , n
• Updates by global best solution x

x2
x1
µ(t) (1 - ?) µ (t - 1) ?x s(t) (1 - ?) s (t
- 1) ?x - µ(t - 1)
11
Extended ACO
ACO (Socha 2004)? complex pheromone
distribution Gaussian kernel PDF
x1
12
Two spirals dataset

• Classification
• 2 classes

13
Training and testing set
14
DACO
Uses sinus units to generate fractal-like solution
15
Results on testing dataset
16
Conclusion

• ACO was able to utilize sinus units
• It was useful to try different methods
• We need to visualize the learning process

17
Thank you