Package org.elasticsearch.search.aggregations.pipeline

Class SimulatedAnealingMinimizer

java.lang.Object

org.elasticsearch.search.aggregations.pipeline.SimulatedAnealingMinimizer

public class SimulatedAnealingMinimizer
extends Object

A cost minimizer which will fit a MovAvgModel to the data. This optimizer uses naive simulated annealing. Random solutions in the problem space are generated, compared against the last period of data, and the least absolute deviation is recorded as a cost. If the new cost is better than the old cost, the new coefficients are chosen. If the new solution is worse, there is a temperature-dependent probability it will be randomly selected anyway. This allows the algo to sample the problem space widely. As iterations progress, the temperature decreases and the algorithm rejects poor solutions more regularly, theoretically honing in on a global minimum.

Constructor Summary

Constructors

Constructor	Description
SimulatedAnealingMinimizer()

Method Summary

Modifier and Type	Method	Description
static MovAvgModel	minimize(MovAvgModel model, EvictingQueue<Double> train, double[] test)	Runs the simulated annealing algorithm and produces a model with new coefficients that, theoretically fit the data better and generalizes to future forecasts without overfitting.

Methods inherited from class java.lang.Object

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

Constructor Details

SimulatedAnealingMinimizer

public SimulatedAnealingMinimizer()

Method Details

minimize

public static MovAvgModel minimize(MovAvgModel model,
 EvictingQueue<Double> train,
 double[] test)

Runs the simulated annealing algorithm and produces a model with new coefficients that, theoretically fit the data better and generalizes to future forecasts without overfitting.

Parameters:

model - The MovAvgModel to be optimized for

train - A training set provided to the model, which predictions will be generated from

test - A test set of data to compare the predictions against and derive a cost for the model

Returns:

A new, minimized model that (theoretically) better fits the data