I extracting features from the signals and I want to classify the signal into 17 classes. In order to achieve the best combination of feature subset, genetic algorithm has been applied for feature selection. For example, I have 10 features. For initialize population, I have random the number and also the position of feature for each population. Then I will applied fitness function to get the accuracy of each population. Next I will pick two parents with the best performance and a crossover will be made on two parents. But, I do not know how to applied mutation? Based on my understanding, the mutation is flip between 0 and 1, but now I have position of 1 to 10. May I know how to perform mutation and how to update the children for each generation?
As you know mutation is an operation that exploits search space by employing property which parents do not have. So, you can apply any of these following two techniques:
1. Traditional mutation
– For each gene, produce a random number. If it is less than a pre-specified threshold, mutate the gene
else, preserve the gene value.
– Swap a pair of genes
2. Non-uniform mutation
Regards,
M M Manjurul Islam
Dear Jing Wei,
For me, there are 2 steps.
Step 1, define the number of features, in can be randomize for each individu. You keep the number of features or not.
Step 2,
If you keep the number of features, turn off the active features (it can be more than 1) and turn on the other features.
If you add the number of features, turn on the inactive feature.
If you reduce the number of features, turn off the active feature.
The other rule maybe appear, so make sure the condition of the individu match for the random condition for step 1.
Best Regards,
Assuming that you have a pool of features and you are going to select the optimal number and order of the features for accurately classifying the signal I would suggest the following methods.
On exceeding the threshold for a mutation,
1. select two positions randomly and swap the corresponding features.
2. select a position randomly and swap the corresponding feature in the chromosome with another feature in your pool.
3. Do the either of above by giving a chance of 50%-50%, or with a bias ( ex. 60%-40%) for one method out of the two, based on the impact (most probably the feature itself may have higher influence than its position ) on the accuracy.
- Badges
- Science topic