GradientDescentSharp
This little library allows to compute a approximate solution for some defined problem with error function, using gradient descent.
Simple example:
//first define a problem
var problem = (IDataAccess<double> x) =>
{
var n = x[0];
//we seek for such value, that n^n=5
var needToMinimize = Math.Pow(n, n) - 5.0;
return Math.Abs(needToMinimize);
};
//then define changing variables
var variables = new ArrayDataAccess<double>(1);
//set variables close to global minima
variables[0] = 1;
//define descent
var descent = new MineDescent(variables, problem)
{
DescentRate = 0.1, // how fast to descent, this value will be adjusted on the fly
Theta = 1e-4, // what precision of found minima we need
DescentRateDecreaseRate = 0.1, // how much decrease DescentRate when we hit a grow of error function
Logger = new ConsoleLogger() // logger for descent progress
};
//do 30 iterations
descent.Descent(30);
System.Console.WriteLine("For problem n^n=5");
System.Console.WriteLine($"Error is {problem(variables)}");
System.Console.WriteLine($"n={variables[0]}");
System.Console.WriteLine($"n^n={Math.Pow(variables[0], variables[0])}");
Output
--------------Mine descent began
Error is 3.8894657589454242
Changed by 0.11053424105457577
-------------
...
-------------
Error is 0.2503619082577506
Changed by 0.7496380917422432
-------------
Error is 0.66669577875009
Changed by 0.4163338704923394
Undo step. Decreasing descentRate.
-------------
...
-------------
Error is 0.00015378896740614323
Changed by 8.779536614600403E-05
--------------Mine done in 24 iterations
For problem n^n=5
Error is 0.00015378896740614323
n=2.1293900000000012
n^n=5.000153788967406
As you can see, when we hit a grow in error function, the descent undo step and decrease descentRate, so we are guaranteed to hit a local minima!
It still is very important to define a good error function and init variables.
Currently there are three providers for gradient descent: adam descent mine descent natural descent
Here is general time-performance measurements for each of them
Descent Score Time ms
adam 290 270
mine 971 162
natural 1739 293
As you can see mine descent currently is fastest one and provides a good results, while natural descent is the most precise
Also I've a good working feed-forward neural network implementation here.
Check out gradient descent playground examples.
Check out neural network playground examples.