package aima.core.search.adversarial;
import java.util.ArrayList;
import java.util.List;
import aima.core.search.framework.Metrics;
/**
* Implements an iterative deepening Minimax search with alpha-beta pruning and
* action ordering. Maximal computation time is specified in seconds. The
* algorithm is implemented as template method and can be configured and tuned
* by subclassing.
*
* @author Ruediger Lunde
*
* @param <STATE>
* Type which is used for states in the game.
* @param <ACTION>
* Type which is used for actions in the game.
* @param <PLAYER>
* Type which is used for players in the game.
*/
public class IterativeDeepeningAlphaBetaSearch<STATE, ACTION, PLAYER> implements AdversarialSearch<STATE, ACTION> {
public final static String METRICS_NODES_EXPANDED = "nodesExpanded";
public final static String METRICS_MAX_DEPTH = "maxDepth";
protected Game<STATE, ACTION, PLAYER> game;
protected double utilMax;
protected double utilMin;
protected int currDepthLimit;
private boolean heuristicEvaluationUsed; // indicates that non-terminal
// nodes
// have been evaluated.
private Timer timer;
private boolean logEnabled;
private Metrics metrics = new Metrics();
/**
* Creates a new search object for a given game.
*
* @param game
* The game.
* @param utilMin
* Utility value of worst state for this player. Supports
* evaluation of non-terminal states and early termination in
* situations with a safe winner.
* @param utilMax
* Utility value of best state for this player. Supports
* evaluation of non-terminal states and early termination in
* situations with a safe winner.
* @param time
* Maximal computation time in seconds.
*/
public static <STATE, ACTION, PLAYER> IterativeDeepeningAlphaBetaSearch<STATE, ACTION, PLAYER> createFor(
Game<STATE, ACTION, PLAYER> game, double utilMin, double utilMax, int time) {
return new IterativeDeepeningAlphaBetaSearch<>(game, utilMin, utilMax, time);
}
/**
* Creates a new search object for a given game.
*
* @param game
* The game.
* @param utilMin
* Utility value of worst state for this player. Supports
* evaluation of non-terminal states and early termination in
* situations with a safe winner.
* @param utilMax
* Utility value of best state for this player. Supports
* evaluation of non-terminal states and early termination in
* situations with a safe winner.
* @param time
* Maximal computation time in seconds.
*/
public IterativeDeepeningAlphaBetaSearch(Game<STATE, ACTION, PLAYER> game, double utilMin, double utilMax,
int time) {
this.game = game;
this.utilMin = utilMin;
this.utilMax = utilMax;
this.timer = new Timer(time);
}
public void setLogEnabled(boolean b) {
logEnabled = b;
}
/**
* Template method controlling the search. It is based on iterative
* deepening and tries to make to a good decision in limited time. Credit
* goes to Behi Monsio who had the idea of ordering actions by utility in
* subsequent depth-limited search runs.
*/
@Override
public ACTION makeDecision(STATE state) {
metrics = new Metrics();
StringBuffer logText = null;
PLAYER player = game.getPlayer(state);
List<ACTION> results = orderActions(state, game.getActions(state), player, 0);
timer.start();
currDepthLimit = 0;
do {
incrementDepthLimit();
if (logEnabled)
logText = new StringBuffer("depth " + currDepthLimit + ": ");
heuristicEvaluationUsed = false;
ActionStore<ACTION> newResults = new ActionStore<>();
for (ACTION action : results) {
double value = minValue(game.getResult(state, action), player, Double.NEGATIVE_INFINITY,
Double.POSITIVE_INFINITY, 1);
if (timer.timeOutOccurred())
break; // exit from action loop
newResults.add(action, value);
if (logEnabled)
logText.append(action).append("->").append(value).append(" ");
}
if (logEnabled)
System.out.println(logText);
if (newResults.size() > 0) {
results = newResults.actions;
if (!timer.timeOutOccurred()) {
if (hasSafeWinner(newResults.utilValues.get(0)))
break; // exit from iterative deepening loop
else if (newResults.size() > 1
&& isSignificantlyBetter(newResults.utilValues.get(0), newResults.utilValues.get(1)))
break; // exit from iterative deepening loop
}
}
} while (!timer.timeOutOccurred() && heuristicEvaluationUsed);
return results.get(0);
}
// returns an utility value
public double maxValue(STATE state, PLAYER player, double alpha, double beta, int depth) {
updateMetrics(depth);
if (game.isTerminal(state) || depth >= currDepthLimit || timer.timeOutOccurred()) {
return eval(state, player);
} else {
double value = Double.NEGATIVE_INFINITY;
for (ACTION action : orderActions(state, game.getActions(state), player, depth)) {
value = Math.max(value, minValue(game.getResult(state, action), //
player, alpha, beta, depth + 1));
if (value >= beta)
return value;
alpha = Math.max(alpha, value);
}
return value;
}
}
// returns an utility value
public double minValue(STATE state, PLAYER player, double alpha, double beta, int depth) {
updateMetrics(depth);
if (game.isTerminal(state) || depth >= currDepthLimit || timer.timeOutOccurred()) {
return eval(state, player);
} else {
double value = Double.POSITIVE_INFINITY;
for (ACTION action : orderActions(state, game.getActions(state), player, depth)) {
value = Math.min(value, maxValue(game.getResult(state, action), //
player, alpha, beta, depth + 1));
if (value <= alpha)
return value;
beta = Math.min(beta, value);
}
return value;
}
}
private void updateMetrics(int depth) {
metrics.incrementInt(METRICS_NODES_EXPANDED);
metrics.set(METRICS_MAX_DEPTH, Math.max(metrics.getInt(METRICS_MAX_DEPTH), depth));
}
/** Returns some statistic data from the last search. */
@Override
public Metrics getMetrics() {
return metrics;
}
/**
* Primitive operation which is called at the beginning of one depth limited
* search step. This implementation increments the current depth limit by
* one.
*/
protected void incrementDepthLimit() {
currDepthLimit++;
}
/**
* Primitive operation which is used to stop iterative deepening search in
* situations where a clear best action exists. This implementation returns
* always false.
*/
protected boolean isSignificantlyBetter(double newUtility, double utility) {
return false;
}
/**
* Primitive operation which is used to stop iterative deepening search in
* situations where a safe winner has been identified. This implementation
* returns true if the given value (for the currently preferred action
* result) is the highest or lowest utility value possible.
*/
protected boolean hasSafeWinner(double resultUtility) {
return resultUtility <= utilMin || resultUtility >= utilMax;
}
/**
* Primitive operation, which estimates the value for (not necessarily
* terminal) states. This implementation returns the utility value for
* terminal states and <code>(utilMin + utilMax) / 2</code> for non-terminal
* states. When overriding, first call the super implementation!
*/
protected double eval(STATE state, PLAYER player) {
if (game.isTerminal(state)) {
return game.getUtility(state, player);
} else {
heuristicEvaluationUsed = true;
return (utilMin + utilMax) / 2;
}
}
/**
* Primitive operation for action ordering. This implementation preserves
* the original order (provided by the game).
*/
public List<ACTION> orderActions(STATE state, List<ACTION> actions, PLAYER player, int depth) {
return actions;
}
///////////////////////////////////////////////////////////////////////////////////////////
// nested helper classes
private static class Timer {
private long duration;
private long startTime;
Timer(int maxSeconds) {
this.duration = 1000 * maxSeconds;
}
void start() {
startTime = System.currentTimeMillis();
}
boolean timeOutOccurred() {
return System.currentTimeMillis() > startTime + duration;
}
}
/** Orders actions by utility. */
private static class ActionStore<ACTION> {
private List<ACTION> actions = new ArrayList<>();
private List<Double> utilValues = new ArrayList<>();
void add(ACTION action, double utilValue) {
int idx = 0;
while (idx < actions.size() && utilValue <= utilValues.get(idx))
idx++;
actions.add(idx, action);
utilValues.add(idx, utilValue);
}
int size() {
return actions.size();
}
}
}