package chase.s2.net; public class SelfOrganizingMap { protected SelfOrganizedNode[] nodes; protected SelfOrganizedNode bmu; protected int[] lattice; protected int statBinIndexes; protected int weights; protected int timeIndex; /** * Creates a self organizing map with a lattice size, stat indexes and * number of node weights as stated. * @param lattice - An array of integers, designating a shape of the lattice * @param statIndexes - number of indexes in output array * @param weights - the nubmer of weights found in a node */ public SelfOrganizingMap(int[] lattice, int statIndexes, int weights) { if(lattice.length < 1) throw new RuntimeException("Lattice requires atleast 1 dimension."); statBinIndexes = statIndexes; this.lattice = lattice; this.weights = weights; int latticeSize = 1; for(int i=0; i<lattice.length; i++) { latticeSize *= lattice[i]; } //end for nodes = new SelfOrganizedNode[latticeSize]; int[] vector = new int[lattice.length]; for(int i=0; i<latticeSize; i++) { nodes[i] = new SelfOrganizedNode(vector, this); initialize(nodes[i], vector); vector[0]++; for(int j=1; j<lattice.length; j++) { if(vector[j-1] >= lattice[j-1]) { vector[j-1] = 0; vector[j]++; } //end if } //end for } //end for } //Notice: In desperate need of a better initialization! private void initialize(SelfOrganizedNode n, int[] position) { //Initializes each node randomly, not very well done for(int i=0; i<weights; i++) n.weights[i] = Math.random(); //I have found that in a 2D map with equal dimensions //that if you set each weight outward from the center //at a distance of about a quarter, and then guassian //it from there with a .5 = dimension/4, that it works //alot better, but my code is far to complex and messy //to include here } public void updateMap(double[] input, int index) { timeIndex++; bmu = getBMU(input); for(int i=0; i<nodes.length;i++) { nodes[i].update(input, index); } } public SelfOrganizedNode getBMU(double[] input) { int bmu = 0; double smallestDifference = Double.POSITIVE_INFINITY; //there is no faster way as the data in the nodes changes constantly for(int i=0; i<nodes.length; i++) { double difference = nodes[i].difference(input); //small shortcut, as we need every ounce of speed if(difference < 0.002) return nodes[i]; if(difference < smallestDifference) { bmu = i; smallestDifference = difference; } } return nodes[bmu]; } } public class SelfOrganizedNode { protected SelfOrganizingMap map; protected int position[]; protected double weights[]; public float statBin[]; protected SelfOrganizedNode(int[] latticePosition, SelfOrganizingMap parent) { position = latticePosition.clone(); map = parent; statBin = new float[map.statBinIndexes]; weights = new double[map.weights]; } public void update(double[] input, int index) { double distance = distance(map.bmu); //Make this smarter, it should degrade over time, and it determines //The size of the area to update, it should be based on the physical //size of the map, not the number of nodes (easiest is the smallest dimension) //I find its best if it slowly decreases in size over about 100 to 1000 ticks/updates double variance = 10/map.timeIndex; double neighborhood = Math.exp(-(variance*variance*distance*distance)/2.0); //This makes the whole thing ALOT faster. Remove for precision or if //you wanna test how you bot handles turn skipping. if(neighborhood < 0.0002) return; //Much more can be done here, this should also be based on time //and it is, however if this is the bmu, then it will get changed //to the input, an undesirable thing later on. I never did this in //my prototype gun, I could of achieved greater precision here for(int i=0; i < map.weights; i++) { weights[i] += neighborhood*(input[i] - weights[i]); } //This is just the index updating, very important to multiply in the neighborhood //otherwise it gets very messy later on, not to much you can do here unlike other areas //.7 is purely a magic number I came up with, I find binIndexes / 30 works well too for(int i=0; i<map.statBinIndexes; i++) { int diff = Math.abs(index - i); float stat = (float)Math.exp(-(.7*.7*diff*diff)/2.0); statBin[i] += neighborhood*stat; // statBin[i] = KTools.rollingAvg(bin[0][i], neighborhood*input, // Math.min(update_num, 5), 100); } } protected float distance(SelfOrganizedNode n) { return distance(position, n.position); } protected double difference(double[] input) { return distance(weights, input); } public static final float distance(int[] p, int[] q) { float k, d = 0; for(int i=0; i<p.length; i++) { d += (k=((float)p[i]-(float)q[i]))*k; } return d; } public static final double distance(double[] p, double[] q) { if(p == null || q == null) return Double.POSITIVE_INFINITY; double k, d = 0; for(int i=0; i<p.length; i++) { d += (k=(p[i]-q[i]))*k; } return d; } }