Since the gradient of the error becomes shallower as the descent nears convergence, this will naturally shrink the updates into the error function's minimum. However, too large of a step size will lead to the descent diverging and too small of a step size will lead to an extremely long descent. Unfortunately, choosing a good step size is a matter of trial and error.
Inheritance: implements MCordingley\Regression\Algorithm\GradientDescent\Schedule\Schedule
Fixed Class Documentation
Example #1
0
 public function testStep()
 {
     $schedule = new Fixed(1.0);
     $schedule->update([1.0]);
     static::assertEquals(1.0, $schedule->step(0));
 }