Genetic algorithms

The sequential calibration has substantial limitations: (i) it cannot automatically choose the pressure levels and temporal windows (hour of the day) for a given meteorological variable, (ii) it cannot handle dependencies between parameters, and (iii) it cannot easily handle new degrees of freedom. For this reason, genetic algorithms (GAs) were implemented in AtmoSwing Optimizer to perform a global optimization of AMs. It allows for the optimization of all parameters jointly in a fully automatic and objective way. See more details in [Horton2017].

Genetic algorithms are powerful but very demanding in terms of computational capacity. They require thousands of assessments to evolve towards the global optimum and thus should be used on a cluster rather than a single computer.

Many options and operator variants are available for genetic algorithms. Based on systematic tests detailed in [Horton2017], some presets were established in order to ease the use of genetic algorithms in AtmoSwing. These presets are listed below, and all options are provided further down.

Options presets

Many options and operator variants control the optimization by genetic algorithms. Recommended configurations were predefined in presets.

 Operators presets

For all presets, the following options are identical: --ga-conv-steps=30 --ga-pop-size=500 --ga-interm-gen=0.5 --ga-ope-nat-sel=0 --ga-ope-coup-sel=2 --ga-ope-cross=7 --ga-cross-bin-pt-nb=2 --ga-cross-bin-share-b=1

The difference between the presets concerns the mutation operator. Providing the following numbers (1-5) to the option ga-config is equivalent to these corresponding presets:

  1. Chromosome of adaptive search radius: --ga-ope-mut=8
  2. Multiscale mutation: --ga-ope-mut=9 --ga-mut-multi-scale-p=0.1
  3. Nonuniform mutation (pmut=0.1, Gmr=50, w=0.1): --ga-ope-mut=4 --ga-mut-non-uni-p=0.1 --ga-mut-non-uni-gens=50 --ga-mut-non-uni-min-r=0.1
  4. Nonuniform mutation (pmut=0.1, Gmr=100, w=0.1): --ga-ope-mut=4 --ga-mut-non-uni-p=0.1 --ga-mut-non-uni-gens=100 --ga-mut-non-uni-min-r=0.1
  5. Nonuniform mutation (pmut=0.2, Gmr=100, w=0.1): --ga-ope-mut=4 --ga-mut-non-uni-p=0.2 --ga-mut-non-uni-gens=100 --ga-mut-non-uni-min-r=0.1

Any of these options can be overridden by specifying it along with ga-config.

All options

The different operators can be controlled with the following options:


Operator choice for natural selection:

  1. ratio elitism
  2. tournament selection

Operator choice for couples selection:

  1. rank pairing
  2. random
  3. roulette wheel on rank
  4. roulette wheel on score
  5. tournament competition

Operator choice for chromosome crossover:

  1. single point crossover
  2. double points crossover
  3. multiple points crossover
  4. uniform crossover
  5. limited blending
  6. linear crossover
  7. heuristic crossover
  8. binary-like crossover
  9. linear interpolation
  10. free interpolation

Operator choice for mutation:

  1. random uniform constant
  2. random uniform variable
  3. random normal constant
  4. random normal variable
  5. non-uniform
  6. self-adaptation rate
  7. self-adaptation radius
  8. self-adaptation rate chromosome
  9. self-adaptation radius chromosome
  10. multi-scale
  11. no mutation
 Size of the population
 Number of generations for convergence
 Ratio of the intermediate generation
 Natural selection - tournament probability
 Couples selection - tournament candidates (2/3)
 Standard crossover - number of points
 Blending crossover - number of points
 Blending crossover - beta shared (1/0)
 Linear crossover - number of points
 Heuristic crossover - number of points
 Heuristic crossover - beta shared (1/0)
 Binary-like crossover - number of points
 Binary-like crossover - beta shared (1/0)
 Uniform mutation - probability
 Normal mutation - probability
 Normal mutation - standard deviation
 Variable uniform mutation - generations nb
 Variable uniform mutation - starting probability
 Variable uniform mutation - end probability
 Variable normal mutation - generations nb for probability
 Variable normal mutation - generations nb for std deviation
 Variable normal mutation - starting probability
 Variable normal mutation - end probability
 Variable normal mutation - starting std deviation
 Variable normal mutation - end std deviation
 Non uniform mutation - probability
 Non uniform mutation - generations nb
 Non uniform mutation - minimum rate
 Multi-scale mutation - probability
[Horton2017](1, 2) Horton, P., Jaboyedoff, M., & Obled, C. (2017). Global Optimization of an Analog Method by Means of Genetic Algorithms. Monthly Weather Review, 145(4), 1275–1294.