Python project to demonstrate the use of generic networks / AI to approximate a function / formula.
Given x and y with a couple of values and a hidden mathematical relation: x * y
data = SampleData()
data.add_values(values={'x': 4.0, 'y': 3.0}, expected=12.0)
data.add_values(values={'x': 27.0, 'y': 2.0}, expected=54.0)
data.add_values(values={'x': 1.0, 'y': 18.0}, expected=18.0)
data.add_values(values={'x': 5.0, 'y': 10.0}, expected=50.0)
data.add_values(values={'x': 12.0, 'y': 4.5}, expected=54.0)
data.add_values(values={'x': 1.5, 'y': 4.5}, expected=6.75)
data.add_values(values={'x': 25.4, 'y': 5.0}, expected=127.0)
data.add_values(values={'x': 20.0, 'y': 17.5}, expected=350.0)
Then the genetic network is created:
# Configure the algorithm:
population_size = 25
genome_length = 1 * 68 + 2 * 68 + 4 * 68 + 8 * 68
ga = GeneticAlgorithm(fitness_function)
ga.generate_binary_population(size=population_size, genome_length=genome_length)
Please note that the size of each gene is 68 (4 for the type and 64 for the value). Since the function is built from a binary tree, each level has twice the number of genes: 1 + 2 + 4 + 8 + ... x 68 genes.
When the network is running for 1000 iterations.
# Run 1000 iteration of the algorithm
ga.run(1000)
The best genome is very likely to return the mathematical relationship:
best_genome, best_fitness = ga.get_best_genome()
best_function = tree_function(best_genome)
print(f'best formula: {formula_function(best_genome)}')
print(f'best fitness: {best_fitness}')
One could even create a diagram:
