[Home]KernelDensity

Robo Home | Changes | Preferences | AllPages

I just created this class to calculate kernel density estimates some time ago, but I didn't get much time now for robocoding, so may be it will be useful for someone.

Given an array of points (pe. a list of bearings) it gives you a good estimation of the probability for a given value (pe. bearing). Probably useful for WaveSurfing or Targeting algorithms.

It has several kernel functions. Just uncomment the one you want to use.

-- Albert

 package apv;

 /**
  * KernelDensity - a class by Albert Perez
  * Based on the following explanation: http://www.xplore-stat.de/tutorials/smoothernode2.html
  * If you are using this code for Robocode, it i released under RWPCL
  * Check http://robowiki.net/cgi-bin/robowiki?RWPCL for the terms.
  */
 public class KernelDensity {
 	
 	public static double densityAtPoint(double[] x, int n, double h, double p, double d) {
 		return densityAtPoint(x, n, h, p, d, n);
 	}
 	
 	public static double densityAtPoint(double[] x, int n, double h, double p, double d, int s) {
 		// x is the array of sample data points (ordered by t, 0 is prev. n-1 is actual
 		// n is the number of existing points in the array	
 		// h is the bandwidth
 		// p is the point at which to calculate the estimate
 		// d is the decay (assumes point at n-1 is the most recent one, then it weights the points as d^t)
 		// decay is just an addition to allow weighting the most recent points 
 		// s is the window (number of points to use for the calculation)
 		if (n==0) return 0;
 		if (s>n) s =n;
 	
 		double d2 = d;
 		double w = 0;
 		double k = 0;
 		for (int i=n-1; i>=n-s; i--) {
 			double u = (x[i]-p) / h;
 			w += d2;
 			
 			//Epanechnikov
 			//k += d2 * (3.0/4.0) * (1 - u*u) * (Math.abs(u)<=1 ? 1 : 0);
 			//Uniform
 			// k += d2 * (1.0/2.0) * (Math.abs(u)<=1 ? 1 : 0);
 			//Triangle
 			// k += d2 * (1- Math.abs(u)) * (Math.abs(u)<=1 ? 1 : 0);
 			//Quartic
 			//k += d2 * (15.0/16.0) * (1 - u*u) * (1 - u*u) * (Math.abs(u)<=1 ? 1 : 0);
 			//Gaussian
 			k += d2 * (1/Math.sqrt(2*Math.PI)) * Math.exp(-0.5 * u * u);
 			//Cosinus
 			//k += d2 * (Math.PI/4.0) * Math.cos(Math.PI/2 * u) * (Math.abs(u)<=1 ? 1 : 0);	
 			
			d2 = d2 * d;
 		}
 		k = k /(w*h);
 		return k;
 	}
	
 	public static double densityAtPoint2D(double[] x, double[] y, int n, double h1, double h2, double px, double py, double d, int s) {
 		// x,y are the arrays of sample data points (ordered by t, 0 is prev. n-1 is actual
 		// n is the number of existing points in the array	
 		// h is the bandwidth
 		// p is the point at which to calculate the estimate
 		// d is the decay (assumes point at n-1 is the most recent one, then it weights the points as d^t)
 		// decay is just an addition to allow weighting the most recent points 
 		// s is the window (number of points to use for the calculation)
 		if (n==0) return 0;
 		if (s>n) s =n;
 	
 		double d2 = d;
 		double w = 0;
 		double k = 0;
 		for (int i=n-1; i>=n-s; i--) {
 			double ux = (x[i]-px) / h1; 
 			double uy = (y[i]-py) / h2;
 			w += d2;
 			
 			//Epanechnikov
 			//k += d2 * (3.0/4.0) * (1 - ux*ux) * (Math.abs(ux)<=1 ? 1 : 0) * (3.0/4.0) * (1 - uy*uy) * (Math.abs(uy)<=1 ? 1 : 0);
 			//Gaussian
 			k += d2 * (1/Math.sqrt(2*Math.PI)) * Math.exp(-0.5 * ux * ux) * (1/Math.sqrt(2*Math.PI)) * Math.exp(-0.5 * uy * uy);
 			
 			d2 = d2 * d;
 		}
 		k = k /(w*h1*h2);
 		return k;
 	}
 	
 public static double derivativeAtPoint(double[] x, int n, double h, double p, double d, int s) {
 		// x is the array of sample data points (ordered by t, 0 is prev. n-1 is actual
 		// n is the number of existing points in the array	
 		// h is the bandwidth
 		// p is the point at which to calculate the estimate
 		// d is the decay (assumes point at n-1 is the most recent one, then it weights the points as d^t)
 		// decay is just an addition to allow weighting the most recent points 
 		// s is the window (number of points to use for the calculation)
 		if (n==0) return 0;
 		if (s>n) s =n;
 	
 		double d2 = d;
 		double w = 0;
 		double k = 0;
 		for (int i=n-1; i>=n-s; i--) {
 			double u = (x[i]-p) / h;
 			w += d2;
 			
 			//Gaussian
 			k += d2 * (1/Math.sqrt(2*Math.PI)) * Math.exp(-0.5 * u * u) * (1/h) * u;
 		
 			d2 = d2 * d;
 		}
 		k = k /(w*h);
 		return k;
 	}
 	
 }

Robo Home | Changes | Preferences | AllPages
Edit text of this page | View other revisions
Last edited July 5, 2005 21:50 EST by Albert (diff)
Search: