Machine Learning Language

The Machine Learning Language helps to train surrogate models. A surrogate model is used in optimisation when the optimisation function is unknown or too expensive to compute. EvoAl offers surrogate learning capabilities that can be configured using the Machine Learning Language for such cases. A simple example can be seen in the following code snippet:

import "definitions" from de.evoal.surrogate.ml;
import "definitions" from de.evoal.surrogate.smile.ml;

import "data" from surrogate;

module training {
  prediction svr
    maps 'x:0'
    to 'y:0'
    using
      layer transfer
        with function 'gaussian-svr'
          mapping 'x:0'
          to 'y:0'
          with parameters
            'ε' := 1.4;
            'σ' := 3.0;
            'soft-margin' := 0.15;
            tolerance := 0.1;

  for _counter in [1 to 10] loop
    predict svr from "data.json"
      and measure
        'cross-validation'(10);
        'R²'();
      end
      and store to "svr_${_counter}.pson"
  end
}

The file consists of two sections. In the first section, a model is configured, and in the second, it is trained and persisted for later use.

Model Configuration

A surrogate model maps i input variables x\text{:}1 \ldots x\text{:}i to j output variables y\text{:}1 \ldots y\text{:}j. The Machine Learning Language allows configuring multiple layered models as a surrogate function. An example can be seen in the following depiction:

graph TD
  x:1{x:1} --> model:1
  x:2{x:2} --> model:1
  x:3{x:3} --> model:1

  x:3 --> model:2
  x:4{x:4} --> model:2
  x:5{x:5} --> model:2

  model:1 --> z:1{z:1}
  model:1 --> z:2{z:2}
  model:2 --> z:3{z:3}

  x:1 --> model:3
  z:1 --> model:3
  z:2 --> model:3
  z:3 --> model:3

  model:3 --> y:1{y:1}

In the example, there are two model layers. The first layer contains model:1 and model:2, and the second layer contains model:3. As shown, while a model can use all input features, it does not have to.

Model Training