@Override
public void ahniEventOccurred(AHNIEvent event) {
if (event.getType() == AHNIEvent.Type.GENERATION_END) {
HyperNEATEvolver evolver = event.getEvolver();
boolean finished = evolver.evolutionFinished();
if ((logChampPerGens >= 0 && finished)
|| (logChampPerGens > 0 && evolver.getGeneration() % logChampPerGens == 0)) {
try {
Chromosome bestPerforming = evolver.getBestPerformingFromLastGen();
NiceWriter outputfile =
new NiceWriter(
new FileWriter(
props.getProperty(HyperNEATConfiguration.OUTPUT_DIR_KEY)
+ "best_performing-"
+ (finished ? "final" : evolver.getGeneration())
+ "-evaluation-"
+ bestPerforming.getId()
+ ".txt"),
"0.00");
Activator substrate = generateSubstrate(bestPerforming, null);
evaluate(bestPerforming, substrate, 0, outputfile);
outputfile.close();
} catch (TranscriberException e) {
logger.info(
"Error transcribing best performing individual: "
+ Arrays.toString(e.getStackTrace()));
} catch (IOException e) {
logger.info("Error opening evaluation log file: " + Arrays.toString(e.getStackTrace()));
}
}
}
}
public void _evaluate(
Chromosome genotype,
Activator substrate,
String baseFileName,
boolean logText,
boolean logImage,
double[] fitnessValues,
Behaviour[] behaviours) {
StackedPoleBalancingExperimentLoader b2dExperiment =
new StackedPoleBalancingExperimentLoader(b2dWorldFilename);
int timestep = _simulate(substrate, maxTimesteps, b2dExperiment);
double f1 = (double) timestep / maxTimesteps;
double fitness = f1;
double perf = 0.0;
// generalization test, if max steps have been reached.
if (timestep == maxTimesteps) {
double score = 0;
for (int i = 0; i < generalizationRuns; i++) {
b2dExperiment = new StackedPoleBalancingExperimentLoader(b2dWorldFilename);
timestep = _simulate(substrate, maxGeneralizationTimesteps, b2dExperiment);
score += timestep;
}
perf = (score / generalizationRuns) / maxGeneralizationTimesteps;
}
if (fitnessValues != null) {
fitnessValues[0] = fitness;
genotype.setPerformanceValue(perf);
}
}
/**
* stores chromosome <code>chrom<code>
*
* @param chrom
*/
private void storeChromosome(Chromosome chrom) {
try {
db.store(chrom);
} catch (Exception e) {
String msg = "PersistenceEventListener: error storing chromosome " + chrom.getId();
logger.error(msg, e);
throw new IllegalStateException(msg + ": " + e);
}
}
private void genotypeEvaluatedEvent(Genotype genotype) {
Collection currentGeneration = genotype.getChromosomes();
// persist generation
if (persistAllChroms || persistLastGen) {
Iterator iter = currentGeneration.iterator();
while (iter.hasNext()) storeChromosome((Chromosome) iter.next());
}
// persist champ
Chromosome c = genotype.getFittestChromosome();
champs.add(c);
if (persistChamps) {
storeChromosome(c);
}
// persist run
try {
db.store(run);
} catch (Exception e) {
logger.error("PersistenceEventListener: error storing run", e);
}
// delete chromosomes we don't want to persist
if (!persistAllChroms) {
previousGeneration.removeAll(currentGeneration);
if (persistChamps) previousGeneration.removeAll(champs);
Iterator it = previousGeneration.iterator();
while (it.hasNext()) {
c = (Chromosome) it.next();
try {
db.deleteChromosome(c.getId().toString());
} catch (Exception e) {
logger.error("error storing chromosome " + c, e);
}
}
}
previousGeneration.clear();
previousGeneration.addAll(currentGeneration);
}
protected int evaluate(
Chromosome genotype, Activator substrate, int evalThreadIndex, NiceWriter logOutput) {
TargetFitnessCalculator.Results results =
fitnessCalculator.evaluate(
substrate,
inputPatterns,
targetOutputPatterns,
minTargetOutputValue,
maxTargetOutputValue,
logOutput);
genotype.setPerformanceValue(results.performance);
return results.fitness;
}
/**
* Create a new neural network from the a genotype.
*
* @param genotype chromosome to transcribe
* @return phenotype If given this will be updated and returned, if NULL then a new network will
* be created.
* @throws TranscriberException
*/
public GridNet newGridNet(Chromosome genotype, GridNet phenotype) throws TranscriberException {
AnjiActivator cppnActivator = cppnTranscriber.transcribe(genotype);
AnjiNet cppn = cppnActivator.getAnjiNet();
int connectionRange = this.connectionRange == -1 ? Integer.MAX_VALUE / 4 : this.connectionRange;
// determine cppn input mapping
// target and source coordinates
int cppnIdxTX = -1, cppnIdxTY = -1, cppnIdxTZ = -1, cppnIdxSX = -1, cppnIdxSY = -1;
// deltas
int cppnIdxDX = -1, cppnIdxDY = -1, cppnIdxAn = -1;
int cppnInputIdx = 1; // 0 is always bias
cppnIdxTX = cppnInputIdx++;
cppnIdxTY = cppnInputIdx++;
cppnIdxSX = cppnInputIdx++;
cppnIdxSY = cppnInputIdx++;
if (depth > 2
&& layerEncodingIsInput) { // if depth == 2 network necessarily feed forward, and only one
// layer of connections can exist
cppnIdxTZ = cppnInputIdx++;
if (!feedForward) {
if (includeDelta) {}
}
}
if (includeDelta) {
cppnIdxDY = cppnInputIdx++; // y delta
cppnIdxDX = cppnInputIdx++; // x delta
}
if (includeAngle) {
cppnIdxAn = cppnInputIdx++; // angle
}
// determine cppn output mapping
int[] cppnIdxW; // weights (either a single output for all layers or one output per layer)
int[] cppnIdxB =
new int[0]; // bias (either a single output for all layers or one output per layer)
int cppnOutputIdx = 0;
if (layerEncodingIsInput) {
cppnIdxW = new int[1];
cppnIdxW[0] = cppnOutputIdx++; // weight value
if (enableBias) {
cppnIdxB = new int[1];
cppnIdxB[0] = cppnOutputIdx++; // bias value
}
} else { // one output per layer
cppnIdxW = new int[depth - 1];
for (int w = 0; w < depth - 1; w++) cppnIdxW[w] = cppnOutputIdx++; // weight value
if (enableBias) {
cppnIdxB = new int[depth - 1];
for (int w = 0; w < depth - 1; w++) cppnIdxB[w] = cppnOutputIdx++; // weight value
}
}
// System.out.println("ii: " + cppnInputIdx + " oi: " + cppnOutputIdx);
double[][][][][][] weights;
double[][][] bias;
boolean createNewPhenotype = (phenotype == null);
if (createNewPhenotype) {
bias = new double[depth - 1][][];
for (int l = 1; l < depth; l++) bias[l - 1] = new double[height[l]][width[l]];
// logger.info("Creating new substrate.");
} else {
bias = phenotype.getBias();
}
double[] cppnInput = new double[cppn.getInputDimension()];
cppnInput[0] = 1; // bias
double cppnTZ, cppnTY, cppnTX, cppnSZ, cppnSY, cppnSX;
if (feedForward) {
if (createNewPhenotype) {
weights = new double[depth - 1][][][][][];
for (int l = 1; l < depth; l++) weights[l - 1] = new double[height[l]][width[l]][1][][];
} else {
weights = phenotype.getWeights();
}
// query CPPN for substrate connection weights
for (int tz = 1; tz < depth; tz++) {
if (depth > 2 && layerEncodingIsInput) {
// double cppnTZ =((double) tz*2) / (depth-1) - 1;
cppnTZ = ((double) tz) / (depth - 1);
cppnInput[cppnIdxTZ] = cppnTZ;
}
for (int ty = 0; ty < height[tz]; ty++) {
// double cppnTY =((double) ty*2) / (height-1) - 1;
if (height[tz] > 1) cppnTY = ((double) ty) / (height[tz] - 1);
else cppnTY = 0.5f;
cppnInput[cppnIdxTY] = cppnTY;
for (int tx = 0; tx < width[tz]; tx++) {
// double cppnTX = ((double) tx*2) / (width-1) - 1;
if (width[tz] > 1) cppnTX = ((double) tx) / (width[tz] - 1);
else cppnTX = 0.5f;
cppnInput[cppnIdxTX] = cppnTX;
// calculate dimensions of this weight target matrix
// (bounded by grid edges)
int dy =
Math.min(height[tz - 1] - 1, ty + connectionRange)
- Math.max(0, ty - connectionRange)
+ 1;
int dx =
Math.min(width[tz - 1] - 1, tx + connectionRange)
- Math.max(0, tx - connectionRange)
+ 1;
// if (createNewPhenotype)
// System.out.println(tz + "," + ty + "," + tx + " dy = " + dy + " dx = " + dx);
if (createNewPhenotype) weights[tz - 1][ty][tx][0] = new double[dy][dx];
double[][] w = weights[tz - 1][ty][tx][0];
// System.out.println("\tsy0 = " + Math.max(0,
// ty-connectionRange) + ", sx0 = " + Math.max(0,
// tx-connectionRange));
// for each connection to zyx
// w{y,x} is index into weight matrix
// s{y,x} is index of source neuron
for (int wy = 0, sy = Math.max(0, ty - connectionRange); wy < dy; wy++, sy++) {
// double cppnSY = ((double) sy * 2) / (height-1) - 1;
if (height[tz - 1] > 1) cppnSY = ((double) sy) / (height[tz - 1] - 1);
else cppnSY = 0.5f;
cppnInput[cppnIdxSY] = cppnSY;
for (int wx = 0, sx = Math.max(0, tx - connectionRange); wx < dx; wx++, sx++) {
// double cppnSX = ((double) sx * 2) / (width-1) - 1;
if (width[tz - 1] > 1) cppnSX = ((double) sx) / (width[tz - 1] - 1);
else cppnSX = 0.5f;
cppnInput[cppnIdxSX] = cppnSX;
// System.out.println(tx + "," + ty + " - " + sx + "," + sy + " (" + cppnTX + "," +
// cppnTY + " - " + cppnSX + "," + cppnSY +
// ")");
// delta
if (includeDelta) {
cppnInput[cppnIdxDY] = cppnSY - cppnTY;
cppnInput[cppnIdxDX] = cppnSX - cppnTX;
}
if (includeAngle) {
double angle = (double) Math.atan2(cppnSY - cppnTY, cppnSX - cppnTX);
angle /= 2 * (double) Math.PI;
if (angle < 0) angle += 1;
cppnInput[cppnIdxAn] = angle;
// System.out.println(tx + "," + ty + " - " + sx + "," + sy + " : " +
// Math.toDegrees(angle*Math.PI*2));
}
cppnActivator.reset();
double[] cppnOutput = cppnActivator.next(cppnInput);
// weights
double weightVal;
if (layerEncodingIsInput)
weightVal =
Math.min(
connectionWeightMax,
Math.max(connectionWeightMin, cppnOutput[cppnIdxW[0]]));
else
weightVal =
Math.min(
connectionWeightMax,
Math.max(connectionWeightMin, cppnOutput[cppnIdxW[tz - 1]]));
if (Math.abs(weightVal) > connectionExprThresh) {
if (weightVal > 0)
weightVal =
(weightVal - connectionExprThresh)
* (connectionWeightMax / (connectionWeightMax - connectionExprThresh));
else
weightVal =
(weightVal + connectionExprThresh)
* (connectionWeightMin / (connectionWeightMin + connectionExprThresh));
w[wy][wx] = weightVal;
} else {
w[wy][wx] = 0;
}
// bias
if (enableBias && sy == ty && sx == tx) {
double biasVal;
if (layerEncodingIsInput)
biasVal =
Math.min(
connectionWeightMax,
Math.max(connectionWeightMin, cppnOutput[cppnIdxB[0]]));
else
biasVal =
Math.min(
connectionWeightMax,
Math.max(connectionWeightMin, cppnOutput[cppnIdxB[tz - 1]]));
if (Math.abs(biasVal) > connectionExprThresh) {
if (biasVal > 0)
biasVal =
(biasVal - connectionExprThresh)
* (connectionWeightMax
/ (connectionWeightMax - connectionExprThresh));
else
biasVal =
(biasVal + connectionExprThresh)
* (connectionWeightMin
/ (connectionWeightMin + connectionExprThresh));
bias[tz - 1][ty][tx] = biasVal;
} else {
bias[tz - 1][ty][tx] = 0;
}
}
// System.out.print("\t" + w[wy][wx]);
}
// System.out.println();
}
// System.out.println();
}
}
// System.out.println();
}
// System.out.println();
// System.out.println();
int[][][] connectionMaxRanges = new int[depth - 1][3][2];
for (int l = 0; l < depth - 1; l++) {
connectionMaxRanges[l][0][0] = -1; // no connections to previous or own layer
connectionMaxRanges[l][0][1] = 1;
connectionMaxRanges[l][1][0] = connectionRange;
connectionMaxRanges[l][1][1] = connectionRange;
connectionMaxRanges[l][2][0] = connectionRange;
connectionMaxRanges[l][2][1] = connectionRange;
}
int[][] layerDimensions = new int[2][depth];
for (int l = 0; l < depth; l++) {
layerDimensions[0][l] = width[l];
layerDimensions[1][l] = height[l];
}
if (createNewPhenotype) {
phenotype =
new GridNet(
connectionMaxRanges,
layerDimensions,
weights,
bias,
activationFunction,
1,
"network " + genotype.getId());
logger.info(
"Substrate has input size "
+ width[0]
+ "x"
+ height[0]
+ " and "
+ phenotype.getConnectionCount(true)
+ " connections.");
} else {
phenotype.setName("network " + genotype.getId());
}
} else { // RECURRENT
if (createNewPhenotype) {
weights = new double[depth - 1][][][][][];
for (int l = 1; l < depth; l++) weights[l - 1] = new double[height[l]][width[l]][][][];
} else {
weights = phenotype.getWeights();
}
// query CPPN for substrate connection weights
for (int tz = 1; tz < depth; tz++) {
cppnTZ = ((double) tz) / (depth - 1);
cppnInput[1] = cppnTZ;
for (int ty = 0; ty < height[tz]; ty++) {
if (height[tz] > 1) cppnTY = ((double) ty) / (height[tz] - 1);
else cppnTY = 0.5f;
cppnInput[2] = cppnTY;
for (int tx = 0; tx < width[tz]; tx++) {
if (width[tz] > 1) cppnTX = ((double) tx) / (width[tz] - 1);
else cppnTX = 0.5f;
cppnInput[3] = cppnTX;
// calculate dimensions of this weight matrix (bounded by grid edges)
int dz =
Math.min(depth - 1, tz + connectionRange)
- Math.max(1, tz - connectionRange)
+ 1; // no connections to input layer
int dy =
Math.min(height[tz] - 1, ty + connectionRange)
- Math.max(0, ty - connectionRange)
+ 1;
int dx =
Math.min(width[tz] - 1, tx + connectionRange)
- Math.max(0, tx - connectionRange)
+ 1;
// System.out.println(z + "," + y + "," + x + " dz = "
// + dz + " dy = " + dy + " dx = " + dx);
weights[tz - 1][ty][tx] = new double[dz][dy][dx];
double[][][] w = weights[tz - 1][ty][tx];
// for each connection to t{zyx}
// w{z,y,x} is index into weight matrix
// s{z,y,x} is index of source neuron
for (int wz = 0, sz = Math.max(1, tz - connectionRange); wz < dz; wz++, sz++) {
cppnSZ = ((double) sz) / (depth - 1);
cppnInput[4] = cppnSZ;
for (int wy = 0, sy = Math.max(0, ty - connectionRange); wy < dy; wy++, sy++) {
if (height[tz] > 1) cppnSY = ((double) sy) / (height[tz] - 1);
else cppnSY = 0.5f;
cppnInput[5] = cppnSY;
for (int wx = 0, sx = Math.max(0, tx - connectionRange); wx < dx; wx++, sx++) {
if (width[tz] > 1) cppnSX = ((double) sx) / (width[tz] - 1);
else cppnSX = 0.5f;
cppnInput[6] = cppnSX;
// delta
if (includeDelta) {
cppnInput[7] = cppnSZ - cppnTZ;
cppnInput[8] = cppnSY - cppnTY;
cppnInput[9] = cppnSX - cppnTX;
}
cppnActivator.reset();
double[] cppnOutput = cppnActivator.next(cppnInput);
// weight
double weightVal =
Math.min(connectionWeightMax, Math.max(connectionWeightMin, cppnOutput[0]));
if (Math.abs(weightVal) > connectionExprThresh) {
if (weightVal > 0)
weightVal =
(weightVal - connectionExprThresh)
* (connectionWeightMax
/ (connectionWeightMax - connectionExprThresh));
else
weightVal =
(weightVal + connectionExprThresh)
* (connectionWeightMin
/ (connectionWeightMin + connectionExprThresh));
w[wz][wy][wx] = weightVal;
} else {
w[wz][wy][wx] = 0;
}
// bias
if (enableBias && wz == 0 && wy == 0 && wx == 0) {
double biasVal =
Math.min(connectionWeightMax, Math.max(connectionWeightMin, cppnOutput[1]));
if (Math.abs(biasVal) > connectionExprThresh) {
if (biasVal > 0)
biasVal =
(biasVal - connectionExprThresh)
* (connectionWeightMax
/ (connectionWeightMax - connectionExprThresh));
else
biasVal =
(biasVal + connectionExprThresh)
* (connectionWeightMin
/ (connectionWeightMin + connectionExprThresh));
bias[tz - 1][ty][tx] = biasVal;
} else {
bias[tz - 1][ty][tx] = 0;
}
}
}
// System.out.println();
}
}
// System.out.println();
}
}
}
if (createNewPhenotype) {
int[][] layerDimensions = new int[2][depth];
for (int l = 0; l < depth; l++) {
layerDimensions[0][l] = width[l];
layerDimensions[1][l] = height[l];
}
phenotype =
new GridNet(
connectionRange,
layerDimensions,
weights,
bias,
activationFunction,
cyclesPerStep,
"network " + genotype.getId());
logger.info("Substrate has " + phenotype.getConnectionCount(true) + " connections.");
} else {
phenotype.setName("network " + genotype.getId());
}
}
return phenotype;
}
