/*
* Here comes the text of your license
* Each line should be prefixed with *
*/
package nars.lab.predict;
import automenta.vivisect.TreeMLData;
import automenta.vivisect.swing.NWindow;
import automenta.vivisect.swing.PCanvas;
import automenta.vivisect.timeline.LineChart;
import automenta.vivisect.timeline.TimelineVis;
import java.awt.Color;
import nars.NAR;
import nars.config.Parameters;
import nars.config.Plugins;
import nars.entity.Concept;
import nars.entity.Task;
import nars.gui.NARSwing;
import nars.gui.output.ConceptsPanel;
import nars.io.TextOutput;
import nars.io.Narsese;
import nars.language.Tense;
import nars.language.Term;
/**
*
* https://github.com/encog/encog-java-examples/blob/master/src/main/java/org/encog/examples/neural/predict/sunspot/PredictSunspotElman.java
* https://github.com/encog/encog-java-examples/blob/master/src/main/java/org/encog/examples/neural/recurrent/elman/ElmanXOR.java
*
* @author me
*/
public class Predict3 {
static double polarize(double v) {
return 2 * (v - 0.5);
}
static double unpolarize(double v) {
v = (v/2) + 0.5;
if (v < 0) v = 0;
if (v > 1.0) v = 1.0;
return v;
}
static float signal = 0;
public static void main(String[] args) throws Narsese.InvalidInputException, InterruptedException {
Parameters.DEBUG = true;
int duration = 8;
float freq = 1.0f / duration * 0.15f;
int minCyclesAhead = 0;
double missingDataRate = 0.1;
double noiseRate = 0.02;
boolean onlyNoticeChange = false;
int thinkInterval = onlyNoticeChange ? 1 : 2;
int discretization = 3;
NAR n = new NAR();
// n.param.duration.set(duration);
//n.param.duration.setLinear(0.5);
//n.param.conceptForgetDurations.set(16);
Discretize discretize = new Discretize(n, discretization);
TreeMLData observed = new TreeMLData("value", Color.WHITE).setRange(0, 1f);
TreeMLData[] predictions = new TreeMLData[discretization];
TreeMLData[] reflections = new TreeMLData[discretization];
for (int i = 0; i < predictions.length; i++) {
predictions[i] = new TreeMLData("Pred" + i,
Color.getHSBColor(0.25f + i / 4f, 0.85f, 0.85f));
//predictions[i].setDefaultValue(0.0);
reflections[i] = new TreeMLData("Refl" + i,
Color.getHSBColor(0.25f + i / 4f, 0.85f, 0.85f));
reflections[i].setDefaultValue(0.0);
//predictions[i].setRange(0, 0.5);
}
TimelineVis tc = new TimelineVis(
new LineChart(observed).thickness(16f).height(128), new LineChart(predictions).thickness(16f).height(128),
new LineChart(reflections).thickness(16f).height(128)
/*new LineChart(predictions[1]).thickness(16f).height(128),
new LineChart(predictions[2]).thickness(16f).height(128),*/
);
//new BarChart(error).height(4)
new NWindow("_", new PCanvas(tc)).show(800, 800, true);
new TextOutput(n, System.out) {
/** dt = relative to center */
public double getPredictionEnvelope(double dt, double duration) {
//guassian curve width=duration
// e^(-(4*x/(dur))^2)
double p = (4 * dt / duration);
return Math.exp( -(p * p) );
}
/**
* only allow future predictions
*/
protected boolean allowTask(Task t) {
if (t.sentence.isEternal()) {
return false;
}
boolean future = false;
if ((t.sentence.getOccurenceTime() > n.time() + minCyclesAhead)) {
System.out.print(n.time() + ".." + t.sentence.getOccurenceTime() + ": ");
future = true;
}
Term term = t.getTerm();
int time = (int) t.sentence.getOccurenceTime();
int value = -1;
float conf = t.sentence.truth.getConfidence();
float expect = 2f * (t.sentence.truth.getFrequency() - 0.5f) * conf;
String ts = term.toString();
if (ts.startsWith("<x_t0 --> y")) {
char cc = ts.charAt("<x_t0 --> y".length());
value = cc - '0';
}
if (value != -1) {
//predictions[(int)value].addPlus(time, expect);
for (int tt = time - duration / 2; tt <= time + duration / 2; tt++) {
double smooth = 1;
expect *= getPredictionEnvelope(time-tt, smooth * duration*2f);
/*
if (future)
predictions[value].addPlus(tt, expect);
else
reflections[value].addPlus(tt, expect);
*/
}
}
return true;
}
};
for (Term t : discretize.getValueTerms("x"))
n.believe(t.toString(), Tense.Present, 0.5f, 0.5f);
//TODO move this to discretize.getDisjunctionBelief
n.believe("<(||,y0,y1,y2,y3,y4,y5,y6,y7) --> y>", Tense.Eternal, 1.0f, 0.95f);
n.run(discretization*4);
//new TextOutput(n, System.out);
Concept[] valueBeliefConcepts = discretize.getValueConcepts("x");
NARSwing.themeInvert();
new NWindow("x", new ConceptsPanel(n, valueBeliefConcepts)).show(900, 600, true);
RNNBeliefPrediction predictor = new RNNBeliefPrediction(n, valueBeliefConcepts) {
@Override
public double[] getTrainedPrediction(double[] input) {
//return new double[] { EngineArray.maxIndex(input) };
return input;
}
@Override
public int getPredictionSize() {
return getInputSize();
//return 1;
}
@Override
protected double[] predict() {
double[] x = super.predict();
if (x == null) return null;
long t = n.time();
for (int i = 0; i < x.length; i++) {
predictions[i].add((int)t, x[i] ); //- x[i*2+1]);
}
return x;
}
};
//new NARSwing(n);
int prevY = -1, curY = -1;
long prevT = n.time();
while (true) {
n.run(thinkInterval);
Thread.sleep(3);
//n.memory.addSimulationTime(1);
signal = (float)Math.max(0, Math.min(1.0, Math.tan(freq * n.time()) * 0.5f + 0.5f));
//signal = (float)Math.sin(freq * n.time()) * 0.5f + 0.5f;
//signal = ((float) Math.sin(freq * n.time()) > 0 ? 1f : -1f) * 0.5f + 0.5f;
signal *= 1.0 + (Math.random()-0.5f)* 2f * noiseRate;
if (Math.random() > missingDataRate)
observed.add((int) n.time(), signal);
prevY = curY;
curY = discretize.i(signal);
if ((curY == prevY) && (onlyNoticeChange)) {
continue;
}
discretize.believe("x", signal, 0);
//input(prevT, Term.get("x_t0"), prevY, Term.get("x_t0"), curY, 1f, n.memory);
//input(prevT, Term.get("x_t0"), prevY, Term.get("x_t0"), 1f-curY, 0.5f, n.memory);
//n.addInput("notice(x_t0," + value + ",(||,y0,y1))!");
//input(prevT, Term.get("x_tMin1"), prevY, Term.get("x_t0"), 1f-y, 0f, n.memory);
//input(Term.get("x_t0"), Term.get(value), 0.0f, 0.0f, n.memory); //input(Term.get("x_t0"), Term.get(otherValue), 1.0f, 0f, n.memory); */
/*
n.believe(xFuncEq0, Tense.Present, 1.0f, y);
n.believe(xFuncEq0, Tense.Present, 0.0f, 1f - y);
n.believe(xFuncEq1, Tense.Present, 1.0f, 1f - y);
n.believe(xFuncEq1, Tense.Present, 0.0f, y);
*/
prevT = n.time();
}
}
}