Information Technology Reference
In-Depth Information
We compare the performance of the BMO and the variants sBMO, cBMO, as
well as the DMO. For the BMO we analyze the influence of the parameter γ .
To get an impression of the influence of the population ratios and the selection
pressure we test a sequence of different settings of population sizes μ and λ .
We compare the performance of the BMO variants to the ES with uncorrelated
Gaussian mutations and N step sizes 2 . We do not compare our self-adaptive
operators to the CMA-ES, since the latter is based on derandomized step size
control.
4.5.1
Unconstrained Real Parameter Optimization
In this section we test the biased mutation on unconstrained real parameter
optimization problems. The test suite consists of the following problems, each
with N = 10 dimensions:
sphere,
double sum,
rosenbrock,
rastrigin,
griewank,
rosenbrock with noise.
The features of the functions are listed in the appendix, see A. The following
experimental settings have been used. The initialization, termination conditions
and the number of runs, i.e. 25 runs, are taken from the recommendations of the
CEC 2005 special session on real-parameter optimization [149]. ES denotes the
uncorrelated Gaussian mutation with N step sizes.
Experimental settings
Population model
(15,100)
1
N
1
2 N
Mutation types
ES, BMO, sBMO, cBMO, DMO, γ =0 . 1, τ 0 =
1 =
Crossover type
intermediate, ρ =2
Selection type
comma
[-100,100], σ (0) =0 . 1
Initialization
Termination
1000 generations, 2000 generations (rosenbrock)
Runs
25
Comparison of Variants
First of all, we test the BMO and its variants on a set of classical test functions.
Table 4.1 shows the experimental results on the sphere function. The results
of the experiments show that the BMO and the variants sBMO and cBMO do
not achieve significant improvements. This also holds for the DMO. It turned
out that the sBMO could not achieve remarkable results within the following
experiments.
2 This standard ES is abbreviated with ES in the tables and figures of this chapter.
 
Search WWH ::




Custom Search