Genetic algorithm learning notes (1)

GeneticsAlgorithm(Genetic Algorithm ).

It can solve any practical problems and implement parallel computing behavior.

The operation object of the genetic algorithm is a set of feasible solutions rather than a single feasible solution. There are multiple search tracks, so it is highly feasible.

Genetic algorithms only need to use the value information of the target, and do not need high-value information such as gradients. Therefore, genetic algorithms are suitable for any large scale, the optimization of highly nonlinear non-continuous multi-peak Functions and the optimization of target functions without resolution expressions has good versatility.

The genetic algorithm merit-based mechanism is a "soft" choice, coupled with good concurrency, making it globally optimized and robust.

The feasible solution of genetic algorithm operations is encoded (such as binary). The objective function is interpreted as the adaptive value of the encoded individual (feasible solution), so it has good operability and simplicity.

Application of Genetic Algorithms in neural networks, fuzzy systems, and machine learning

Neural Networks:

I: network connection weight optimization

II: Network Topology Optimization

The biggest weakness of the reverse propagation (BP) Method Based on the Gradient Descent Method: local problem minimization and inability to learn the network topology.

 

Terms of genetic algorithm:

1. Chromosome: Chromosome
2. Compatible score: Adaptive score
3. GENE: Gene
4. Genome: Genome
5. EPOCH: Times
6. Convergence: Convergence
7. Cross: crossbreeding
8. Variation: mutation
9. Iteration: Iteration

Convergence (convergence)

Gambling wheel Algorithm

If the population converges to a special type of chromosome too quickly and the mutation rate is very low, a solution cannot be found by its own variations. Because of the gambling round selection method, the most suitable chromosomes of any generation in the age cannot be uploaded to the next generation.

 

Genetic Operation:

Reproduction (copy ):

The process of selecting a strong individual string from an old population to generate a new population.

The copy operation determines whether an individual is eliminated based on the individual's fitness (Merits). This is just a probability, indicating the size of the opportunity, not the absolute value.

Crossover (crossover ):

Select two fathers to exchange some of their codes.

Simulate the breeding process, select two individuals in the matching pool, randomly select the switching point and swap the part after the switching point.

Mutation ):

A simple operation is to change the number at a certain position on the digital string.

The variation randomly changes the genetic gene at a low probability.

 

Parameter Selection:

The selection of these parameters has a major impact, but there is no quick and effective rule, just some suggestions. Crossover probability: 0.6 ~ 0.95 variation rate: 0.001 ~ 0.01

The total number of chromosomes in a population is called the population scale. This has a significant impact on algorithms. Being too small is not conducive to evolution, but too large is not conducive to computing. There are different scales for different problems. One useful principle is that the number of chromosomes is two times the length of chromosomes.

 

Genetic algorithm selection method:

I: Select roulette wheel selection

This is a proportional choice. Due to the random operation, the error is large, and sometimes individuals with high fitness cannot choose

II: Select tournament selection for the championship

Each time, several individuals with the highest fitness are selected to inherit from the next generation. The advantage is that there is no requirement for positive and negative values for individual fitness, but this method is more random and has a greater random error. However, there is a high probability that the optimal individual will be selected and the worst individual will be eliminated.

III: Random traversal Selection

IV: Selection Based on Population Communication

Problems that often occur in genetic algorithms may lead to premature convergence or super individuals.

 

Roulette wheel selection:

This is a proportional choice.