In this project, I implemented and trained a network to do regression, binary classification and multi class classification. I used evolutionary strategy for training.

In this implementation, the length of chromosome is not fixed, and the algorithm finds the optimum one(find the optimum number of bases).

List of contents:

• Evolutionary Strategy (ES) Parameters
• Network Architecture
• How to use
• Some Results

If your data is two dimensional:

<σ,x1,y1,r1,x2,y2,r2,...>

I select them randomly and uniformly.

τ is changed by random normal distribution.

σ’ = σ • exp(τ • N(0,1))
x’i = xi + σ’ • N(0,1)

Note that τ α 1/ n½

It just produce one child.

I assign Pc=0.4

I use μ+λ and q-tournament methods. I take λ≈7μ and q=5.

• Regression:

e = 0.5 . Transpose(y-y*) . (y-y*)    fitness = 1/e

• Binary classifier:

fitness = 1-sum(abs(sign(y)-y*))/2L

• Multi classifier:

fitness = 1-sum(sign(abs(indexMax(y)-indexMax(y*)))))/L

I use one-hot encoding in multi classification.

You can see the network architecture for regression below:

You can see the network architecture for multi class classifier below:

You can see how I used it in run.py.

You can feed xlsx data by read_excel method.

The result of the dataset in dataset folder, is shown below:

• Regression:

error: 7%

• Classification:

precision: 95%

• Binary Classification:

precision: 98%