moga
Multi-objective Genetic Algorithm (a.k.a Evolutionary Algorithm)
Topics
package_jega
Specification
Alias: None
Arguments: None
Child Keywords:
Required/Optional |
Description of Group |
Dakota Keyword |
Dakota Keyword Description |
---|---|---|---|
Optional |
Select the fitness type for JEGA methods |
||
Optional |
Select a replacement type for JEGA methods |
||
Optional |
Specify the type of niching pressure |
||
Optional |
Select the convergence type for JEGA methods |
||
Optional |
Post process the final solution from
|
||
Optional |
Number of iterations allowed for optimizers and adaptive UQ methods |
||
Optional |
Number of function evaluations allowed for optimizers |
||
Optional |
Turn on scaling for variables, responses, and constraints |
||
Optional |
Set the initial population size in JEGA methods |
||
Optional |
Specify the name of a log file |
||
Optional |
Print every population to a population file |
||
Optional |
Specify how to initialize the population |
||
Optional |
Select a crossover type for JEGA methods |
||
Optional |
Select a mutation type for JEGA methods |
||
Optional |
Seed of the random number generator |
||
Optional |
Stopping criterion based on objective function or statistics convergence |
||
Optional |
Identifier for model block to be used by a method |
Description
moga
stands for Multi-objective Genetic Algorithm, which is a
global optimization method that does Pareto optimization for multiple
objectives. It supports general constraints and a mixture of real and
discrete variables.
Constraints
moga
can utilize linear constraints using the keywords:
-* linear_inequality_constraint_matrix -* linear_inequality_lower_bounds -* linear_inequality_upper_bounds -* linear_inequality_scale_types -* linear_inequality_scales -* linear_equality_constraint_matrix -* linear_equality_targets -* linear_equality_scale_types -* linear_equality_scales
Configuration
The genetic algorithm configurations are:
-fitness -replacement -niching -convergence -postprocessor -initialization -crossover -mutation -population size
The steps followed by the algorithm are listed below. The configurations will effect how the algorithm completes each step.
Stopping Criteria
The moga
method respects the max_iterations
and
max_function_evaluations
method independent controls to provide
integer limits for the maximum number of generations and function
evaluations, respectively.
The algorithm also stops when convergence is reached. This involves repeated assessment of the algorithm’s progress in solving the problem, until some criterion is met.
The specification for convergence in a moga can either be
metric_tracker
or can be omitted all together. If omitted, no
convergence algorithm will be used and the algorithm will rely on
stopping criteria only.
Expected Outputs
The moga
method respects the output
method independent control
to vary the amount of information presented to the user during
execution.
The final results are written to the Dakota tabular output. Additional
information is also available - see the log_file
and
print_each_pop
keywords.
Note that moga and SOGA create additional output files during execution. “finaldata.dat” is a file that holds the final set of Pareto optimal solutions after any post-processing is complete. “discards.dat” holds solutions that were discarded from the population during the course of evolution.
It can often be useful to plot objective function values from these files to visually see the Pareto front and ensure that finaldata.dat solutions dominate discards.dat solutions. The solutions are written to these output files in the format “Input1…InputN..Output1…OutputM”.
Expected HDF5 Output
If Dakota was built with HDF5 support and run with the
hdf5
keyword, this method
writes the following results to HDF5:
Best Objective Functions (when
objective_functions
) are specified)Calibration (when
calibration_terms
are specified)
Important Notes
The pool of potential members is the current population and the current set of offspring.
Choice of fitness assessors is strongly related to the type of replacement algorithm being used and can have a profound effect on the solutions selected for the next generation.
If using the fitness types layer_rank
or domination_count
, it is
strongly recommended that you use the replacement_type
below_limit
(although the roulette wheel selectors can also be used).
The functionality of the domination_count selector of JEGA v1.0 can
now be achieved using the domination_count
fitness type and
below_limit
replacement type.
Theory
The basic steps of the moga
algorithm are as follows:
Initialize the population
Evaluate the population (calculate the values of the objective function and constraints for each population member)
Loop until converged, or stopping criteria reached
Perform crossover
Perform mutation
Evaluate the new population
Assess the fitness of each member in the population
Replace the population with members selected to continue in the next generation
Apply niche pressure to the population
Test for convergence
Perform post processing
If moga is used in a hybrid optimization method (which requires one optimal solution from each individual optimization method to be passed to the subsequent optimization method as its starting point), the solution in the Pareto set closest to the “utopia” point is given as the best solution. This solution is also reported in the Dakota output.
This “best” solution in the Pareto set has minimum distance from the utopia point. The utopia point is defined as the point of extreme (best) values for each objective function. For example, if the Pareto front is bounded by (1,100) and (90,2), then (1,2) is the utopia point. There will be a point in the Pareto set that has minimum L2-norm distance to this point, for example (10,10) may be such a point.
If moga is used in a method which may require passing multiple
solutions to the next level (such as the surrogate_based_global
method or hybrid
methods), the orthogonal_distance
postprocessor type may be used to specify the distances between each
solution value to winnow down the solutions in the full Pareto front
to a subset which will be passed to the next iteration.