hex.optimization

Class L_BFGS

java.lang.Object

hex.optimization.L_BFGS

public class L_BFGS
extends java.lang.Object

Created by tomasnykodym on 9/15/14. L-BFGS optmizer implementation. Use by calling solve() and passing in your own gradient computation function.

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	L_BFGS.GradientInfo
static class	L_BFGS.GradientSolver To be overriden to provide gradient computation specific for given problem.
static class	L_BFGS.History Keeps L-BFGS history ie curvature information recorded over the last m steps.
static class	L_BFGS.L_BFGS_Params Internal parameters affecting behavior of L-BFGS solver.
static class	L_BFGS.Result

Field Summary

Fields

Modifier and Type	Field and Description
static double	c1

Constructor Summary

Constructors

Constructor and Description
L_BFGS()

Method Summary

Methods

Modifier and Type	Method and Description
static L_BFGS.Result	solve(L_BFGS.GradientSolver gslvr, L_BFGS.L_BFGS_Params params, double[] coefs) Solve the optimization problem defined by the user-supplied gradient function using L-BFGS algorithm.
static L_BFGS.Result	solve(L_BFGS.GradientSolver gslvr, L_BFGS.L_BFGS_Params params, L_BFGS.History hist, double[] coefs) Solve the optimization problem defined by the user-supplied gradient function using L-BFGS algorithm.
static double[]	startCoefs(int n, long seed)

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Field Detail

c1

public static final double c1

See Also:

Constant Field Values

Constructor Detail

L_BFGS

public L_BFGS()

Method Detail

solve

public static final L_BFGS.Result solve(L_BFGS.GradientSolver gslvr,
                  L_BFGS.L_BFGS_Params params,
                  double[] coefs)

Solve the optimization problem defined by the user-supplied gradient function using L-BFGS algorithm. Will result into multiple (10s to 100s or even 1000s) calls of the user-provided gradient function. Outside of that it does only limited single threaded computation (order of number of coefficients). The gradient is likely to be the most expensive part and key for good perfomance.

Parameters:

gslvr - - user gradient function

params - - internal L-BFGS parameters.

Returns:

Optimal solution (coefficients) + gradient info returned by the user gradient function evaluated at the found optmimum.

solve

public static final L_BFGS.Result solve(L_BFGS.GradientSolver gslvr,
                  L_BFGS.L_BFGS_Params params,
                  L_BFGS.History hist,
                  double[] coefs)

Solve the optimization problem defined by the user-supplied gradient function using L-BFGS algorithm. Will result into multiple (10s to 100s or even 1000s) calls of the user-provided gradient function. Outside of that it does only limited single threaded computation (order of number of coefficients). The gradient is likely to be the most expensive part and key for good perfomance.

Parameters:

gslvr - - user gradient function

hist - - history of computation

coefs - - starting solution

Returns:

Optimal solution (coefficients) + gradient info returned by the user gradient function evaluated at the found optmimum.

startCoefs

public static double[] startCoefs(int n,
                  long seed)