This repo is an approach to solving the known problem of regression using a evolutionary algorithm. More specifically using evolution strategy.
The approximation of the function is done by a sum of Gaussian kernels:
where $ K_{G_i} $ is defined as:
Then what we are trying to optimize is the set of parameters:
To minimize the error of the pair:
The following functions have been added as a sample:
| Funcitons collection | |
|---|---|
Function 1:  |
Function 2:  |
Function 3:  |
Function 4:  |
Function 5:  |
More functions can be added |
We are using the following error:
However mean squared error and absolute mean error are available in the code as well.
| Predicted function vs True function | Mean and max fitness per generation |
|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
Since the error we are using allows to track a pseudo hit ratio, here are the outputs of such metric for all 5 functions:
| Function | Hit ratios for fittest |
|---|---|
| Function1 |  |
| Function2 |  |
| Function3 |  |
| Function4 |  |
| Function5 |  |
The file algo_params.json contains the parametrization of the algorithm.
{
"num_runs": 10,
"num_generations": 40,
"population_size": 41,
"n_kernels": 4,
"σ_initial": 1,
"λ_children": 200,
"μ_parents": 30,
"recombination_type": "discrete_recombination",
"mutation_type": "uncorr_mutation_n_stepsize",
"error_function": "weighted_mean_abs_error",
"threshold": 0.1,
"lower_weight": 1.0,
"upper_weight": 10.0,
"evaluate_function": "function1",
"x_range": [-1, 3],
"survivor_selection": "all"
"output_file": ""
}
Change the parameters desired and then run algo_runner.jl.
Error, functions and hit ratio graphs will be generated under the folder outputs/ and then given the name in the parameters as output_file. If empty or blank spaces, graphs wont be saved and will be displayed instead.