public boolean[] getAction() {
// Take random action with some probability
if (numGenerator.nextFloat() < getEpsilonGreedy()) {
int randIndex = numGenerator.nextInt(possibleActions.size());
action = possibleActions.get(randIndex);
}
// Otherwise return best action (calculated in integrateObservation)
return action;
}
private boolean[] findBestAction(Environment env, int[] state) {
ArrayList<boolean[]> allActions = getPossibleActions(env);
// Calculate score for all possible next actions
double[] qscores = new double[allActions.size()];
for (int i = 0; i < allActions.size(); i++) {
boolean[] action = allActions.get(i);
StateActionPair sap = new StateActionPair(state, action);
qscores[i] = evalScore(sap);
}
// Find ArgMax over all actions using calculated scores
int ind = numGenerator.nextInt(allActions.size());
boolean[] bestAction = allActions.get(ind);
bestScore = qscores[ind];
for (int i = 1; i < allActions.size(); i++) {
if (qscores[i] > bestScore) {
bestScore = qscores[i];
bestAction = allActions.get(i);
}
}
// System.out.println(bestScore);
return (bestAction);
}
private double[][] extractFeatures(StateActionPair sap) {
// Feature extractor
int[] state = sap.getState();
double[][] features = new double[1][numFeatures];
int ind = 0;
for (int i = 0; i < state.length; i++) {
features[0][ind] = (state[i] == 0) ? 0.0 : 1.0;
ind++;
}
for (int i = 0; i < possibleActions.size(); i++) {
if (Arrays.equals(possibleActions.get(i), sap.getAction())) {
features[0][ind] = 1.0;
break;
}
ind++;
}
// Bias term
features[0][numFeatures - 1] = 1.0;
return (features);
}