Monday, October 17, 2005

The 300 Dimensions Sphere Model

Hallo, Friends!!

I wrote here an Evolution Strategy, (mu+lambda)-ES, to solve the function called "The Sphere Model". Well, this function is not so difficult to solve and is often applied to verify the convergence rate of some Evolutionary Algorithms, like Evolution Strategies.

The Sphere Model is as follow:

F = sum(x(i)^2)
i = 1:n

As you can see the function is not so complicated, but this does not shrink its merit inside the Evolutionary Computation World and, as I said before, that function is very useful to analyse some properties of the Evolutionary Algorithms.

The Sphere Model Configurations:

Kind of Problem: Minimization
Number of Dimensions(n or variables): 300
Variables' Search Space: x(i) = [-100:100], i = 1:300
Global Optimum: F = 0.0
Optima Variables: x(i) = 0.0, i = 1:300

Below you can see a two dimensional Sphere Model:



Another view:





To solve the 300 dimensions Sphere Model I used an Evolution Strategy, its configuration is:

Number of Parents(mu): 15
Number of Offspring(lambda): 35
Number of Generations: 100,000
Mutation: Yes
Crossover: Yes
Selection: Plus - (mu + lambda)-ES


Best Initial Fitness: 920084.13354163163
Best Final Fitness(after 100,000 generations): 1.5747132969144893e-127
Simulation Time: 12,243 s

Below you can see a chart that show you the progress of the best fitness versus the number of generations.



You can note that the Evolution Strategy goes close to the optimum very "fast" and the major part of the time is spent on fine tunning the best solution(s).

Another chart:




Well, I do not know the value of each of the 300 variables (because I did not take them), but I can guarantee you that all 300 variables were very close to the optimum, a prove for this is the value of the best final fitness: 1.5747132969144893e-127

Até Mais!! (Until Later!!)

Nosophorus
Charles Darwin Has A Posse Check Google Page Rank