wfolsom/msproj
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Bug in TDL seems to be fixed. Issue was incorrect calculation of next state reward (reuse of current state) in epsilon-greedy learner.

Browse Source
This commit is contained in:
Woody Folsom
2012-11-28 19:45:03 -05:00
parent 214bdcd032
commit d24e7aee97
6 changed files with 130 additions and 101 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
src/net/woodyfolsom/msproj/ann/Connection.java
View File

@@ -7,7 +7,7 @@ public class Connection {
private int src;
private int dest;
private double weight;
private transient double lastDelta = 0.0;
//private transient double lastDelta = 0.0;
private transient double trace = 0.0;
public Connection() {
@@ -23,7 +23,7 @@ public class Connection {
public void addDelta(double delta) {
this.trace = delta;
this.weight += delta;
this.lastDelta = delta;
//this.lastDelta = delta;
}
@XmlAttribute
@@ -31,10 +31,10 @@ public class Connection {
return dest;
}
@XmlTransient
public double getLastDelta() {
return lastDelta;
}
//@XmlTransient
//public double getLastDelta() {
// return lastDelta;
//}
@XmlAttribute
public int getSrc() {
@@ -100,6 +100,6 @@ public class Connection {
@Override
public String toString() {
return "Connection(" + src + ", " + dest +"), weight: " + weight;
return "Connection(src: " + src + ",dest: " + dest + ", trace:" + trace +"), weight: " + weight;
}
}

5
src/net/woodyfolsom/msproj/ann/NNDataPair.java
View File

@@ -16,4 +16,9 @@ public class NNDataPair {
public NNData getIdeal() {
return ideal;
}
@Override
public String toString() {
return input.toString() + " => " + ideal.toString();
}
}

106
src/net/woodyfolsom/msproj/ann/TTTFilterTrainer.java
View File

@@ -1,8 +1,6 @@
package net.woodyfolsom.msproj.ann;
import java.io.File;
import java.io.FileNotFoundException;
import java.io.FileOutputStream;
import java.util.ArrayList;
import java.util.List;
@@ -15,49 +13,74 @@ import net.woodyfolsom.msproj.tictactoe.Policy;
import net.woodyfolsom.msproj.tictactoe.RandomPolicy;
import net.woodyfolsom.msproj.tictactoe.State;
public class TTTFilterTrainer { //implements epsilon-greedy trainer? online version of NeuralNetFilter
public class TTTFilterTrainer { // implements epsilon-greedy trainer? online
// version of NeuralNetFilter
public static void main(String[] args) throws FileNotFoundException {
double alpha = 0.0;
double lambda = 0.9;
int maxGames = 15000;
double alpha = 0.15;
double lambda = .95;
int maxGames = 1000;
new TTTFilterTrainer().trainNetwork(alpha, lambda, maxGames);
}
public void trainNetwork(double alpha, double lambda, int maxGames) throws FileNotFoundException {
///
FeedforwardNetwork neuralNetwork = new MultiLayerPerceptron(true, 9,5,1);
public void trainNetwork(double alpha, double lambda, int maxGames)
throws FileNotFoundException {
FeedforwardNetwork neuralNetwork = new MultiLayerPerceptron(true, 9, 6,
1);
neuralNetwork.setName("TicTacToe");
neuralNetwork.initWeights();
TrainingMethod trainer = new TemporalDifference(0.5,0.5);
TrainingMethod trainer = new TemporalDifference(alpha, lambda);
System.out.println("Playing untrained games.");
for (int i = 0; i < 10; i++) {
System.out.println("" + (i+1) + ". " + playOptimal(neuralNetwork).getResult());
System.out.println("" + (i + 1) + ". "
+ playOptimal(neuralNetwork).getResult());
}
System.out.println("Learning from " + maxGames + " games of random self-play");
System.out.println("Learning from " + maxGames
+ " games of random self-play");
int gamesPlayed = 0;
List<RESULT> results = new ArrayList<RESULT>();
do {
GameRecord gameRecord = playEpsilonGreedy(0.90, neuralNetwork, trainer);
GameRecord gameRecord = playEpsilonGreedy(0.90, neuralNetwork,
trainer);
System.out.println("Winner: " + gameRecord.getResult());
gamesPlayed++;
results.add(gameRecord.getResult());
} while (gamesPlayed < maxGames);
///
System.out.println("Learned network after " + maxGames + " training games.");
double[][] validationSet = new double[8][];
System.out.println("Learned network after " + maxGames
+ " training games.");
for (int i = 0; i < results.size(); i++) {
if (i % 10 == 0) {
System.out.println("" + (i+1) + ". " + results.get(i));
System.out.println("" + (i + 1) + ". " + results.get(i));
}
}
evalTestCases(neuralNetwork);
System.out.println("Playing optimal games.");
for (int i = 0; i < 10; i++) {
System.out.println("" + (i + 1) + ". "
+ playOptimal(neuralNetwork).getResult());
}
/*
* File output = new File("ttt.net");
*
* FileOutputStream fos = new FileOutputStream(output);
*
* neuralNetwork.save(fos);
*/
}
private void evalTestCases(FeedforwardNetwork neuralNetwork) {
double[][] validationSet = new double[8][];
// empty board
validationSet[0] = new double[] { 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0 };
@@ -81,10 +104,8 @@ public class TTTFilterTrainer { //implements epsilon-greedy trainer? online vers
0.0, -1.0 };
// about to win
validationSet[7] = new double[] {
-1.0, 1.0, 1.0,
1.0, -1.0, 1.0,
-1.0, -1.0, 0.0 };
validationSet[7] = new double[] { -1.0, 1.0, 1.0, 1.0, -1.0, 1.0, -1.0,
-1.0, 0.0 };
String[] inputNames = new String[] { "00", "01", "02", "10", "11",
"12", "20", "21", "22" };
@@ -92,18 +113,6 @@ public class TTTFilterTrainer { //implements epsilon-greedy trainer? online vers
System.out.println("Output from eval set (learned network):");
testNetwork(neuralNetwork, validationSet, inputNames, outputNames);
System.out.println("Playing optimal games.");
for (int i = 0; i < 10; i++) {
System.out.println("" + (i+1) + ". " + playOptimal(neuralNetwork).getResult());
}
/*
File output = new File("ttt.net");
FileOutputStream fos = new FileOutputStream(output);
neuralNetwork.save(fos);*/
}
private GameRecord playOptimal(FeedforwardNetwork neuralNetwork) {
@@ -113,6 +122,8 @@ public class TTTFilterTrainer { //implements epsilon-greedy trainer? online vers
State state = gameRecord.getState();
System.out.println("Playing optimal game:");
do {
Action action;
State nextState;
@@ -120,23 +131,25 @@ public class TTTFilterTrainer { //implements epsilon-greedy trainer? online vers
action = neuralNetPolicy.getAction(gameRecord.getState());
nextState = gameRecord.apply(action);
//System.out.println("Action " + action + " selected by policy " + selectedPolicy.getName());
//System.out.println("Next board state: " + nextState);
System.out.println("Action " + action + " selected by policy " +
neuralNetPolicy.getName());
System.out.println("Next board state: " + nextState);
state = nextState;
} while (!state.isTerminal());
//finally, reinforce the actual reward
// finally, reinforce the actual reward
return gameRecord;
}
private GameRecord playEpsilonGreedy(double epsilon, FeedforwardNetwork neuralNetwork, TrainingMethod trainer) {
private GameRecord playEpsilonGreedy(double epsilon,
FeedforwardNetwork neuralNetwork, TrainingMethod trainer) {
GameRecord gameRecord = new GameRecord();
Policy randomPolicy = new RandomPolicy();
Policy neuralNetPolicy = new NeuralNetPolicy(neuralNetwork);
//System.out.println("Playing epsilon-greedy game.");
// System.out.println("Playing epsilon-greedy game.");
State state = gameRecord.getState();
NNDataPair statePair;
@@ -158,17 +171,20 @@ public class TTTFilterTrainer { //implements epsilon-greedy trainer? online vers
nextState = gameRecord.apply(action);
statePair = NNDataSetFactory.createDataPair(state);
NNDataPair nextStatePair = NNDataSetFactory.createDataPair(nextState);
trainer.iteratePattern(neuralNetwork, statePair, nextStatePair.getIdeal());
NNDataPair nextStatePair = NNDataSetFactory
.createDataPair(nextState);
trainer.iteratePattern(neuralNetwork, statePair,
nextStatePair.getIdeal());
}
//System.out.println("Action " + action + " selected by policy " + selectedPolicy.getName());
// System.out.println("Action " + action + " selected by policy " +
// selectedPolicy.getName());
//System.out.println("Next board state: " + nextState);
// System.out.println("Next board state: " + nextState);
state = nextState;
} while (!state.isTerminal());
//finally, reinforce the actual reward
// finally, reinforce the actual reward
statePair = NNDataSetFactory.createDataPair(state);
trainer.iteratePattern(neuralNetwork, statePair, statePair.getIdeal());
@@ -180,7 +196,7 @@ public class TTTFilterTrainer { //implements epsilon-greedy trainer? online vers
for (int valIndex = 0; valIndex < validationSet.length; valIndex++) {
NNDataPair dp = new NNDataPair(new NNData(inputNames,
validationSet[valIndex]), new NNData(outputNames,
validationSet[valIndex]));
new double[] {0.0}));
System.out.println(dp + " => " + neuralNetwork.compute(dp));
}
}

9
src/net/woodyfolsom/msproj/ann/TemporalDifference.java
View File

@@ -83,7 +83,7 @@ public class TemporalDifference extends TrainingMethod {
private void updateWeights(FeedforwardNetwork neuralNetwork, double predictionError) {
for (Connection connection : neuralNetwork.getConnections()) {
Neuron srcNeuron = neuralNetwork.getNeuron(connection.getSrc());
/*Neuron srcNeuron = neuralNetwork.getNeuron(connection.getSrc());
Neuron destNeuron = neuralNetwork.getNeuron(connection.getDest());
double delta = alpha * srcNeuron.getOutput()
@@ -91,6 +91,13 @@ public class TemporalDifference extends TrainingMethod {
// TODO allow for momentum
// double lastDelta = connection.getLastDelta();
connection.addDelta(delta);*/
Neuron srcNeuron = neuralNetwork.getNeuron(connection.getSrc());
Neuron destNeuron = neuralNetwork.getNeuron(connection.getDest());
double delta = alpha * srcNeuron.getOutput()
* destNeuron.getGradient() + connection.getTrace() * lambda;
//TODO allow for momentum
//double lastDelta = connection.getLastDelta();
connection.addDelta(delta);
}
}

3
src/net/woodyfolsom/msproj/tictactoe/NeuralNetPolicy.java
View File

@@ -24,7 +24,8 @@ public class NeuralNetPolicy extends Policy {
for (Action action : validMoves) {
State nextState = state.apply(action);
NNDataPair dataPair = NNDataSetFactory.createDataPair(state);
//NNDataPair dataPair = NNDataSetFactory.createDataPair(state);
NNDataPair dataPair = NNDataSetFactory.createDataPair(nextState);
//estimated reward for X
scores.put(action, neuralNet.compute(dataPair).getValues()[0]);
}