A javascript library implementing useful multivariate function optimization procedures, which allow to find a local minimum of some function of a vector argument. Such argument is a javascript array.
Example zero order optimization in JavaScript (no need to provide gradient information!):
<script src="./optimization.js"></script>
<script>
// objective that needs to be minimized
fnc = function (v) {
var result = 0.0;
for (var i = 0; i < v.length; i++){
result = result + v[i] * v[i]
}
return result;
};
var x0 = [1.0, -1.0, 0.5, -0.5, 0.25, -0.25]; // a somewhat random initial vector
// Powell method can be applied to zero order unconstrained optimization
var solution = optimjs.minimize_Powell(fnc, x0);
</script>
For more examples, check out examples folder.
For use in browser, use this:
<script src="https://unpkg.com/optimization-js@latest/dist/optimization.js"></script>
In order to use this in node, use npm:
npm install optimization-jsDocumentation is hosted on github pages here: http://optimization-js.github.io/optimization-js/.
Gradient free:
- Genetic optimization algorithms Useful when your function takes as input arguments of mixed type, such as categorical and numerical values.
- Variation of Powell zero order minimization method. Very useful method for prototyping. Typically works decently for problems with 100 - 1000 variables (vector argument size).
Requires gradient:
- Limited memory Broyden–Fletcher–Goldfarb–Shanno method (L-BFGS). Very popular and powerful minimization algorithm. Uses approximation to the Hessian based on recorded gradients over the last m function evaluations. Involves numerical division, and because of this can be unstable, so use at your own risk.
- Vanilla gradient descent. Performs gradient descent using user provided function and its gradient. Can be used instead of L-BFGS in case the latter is unstable on your problem.
See JavaScript files in examples folder for examples on how to use the algorithms.