Add Monte Carlo Tree Search Algorithm (#2588)

This commit is contained in:
Aitor Fidalgo Sánchez 2021-10-16 17:12:33 +03:00 committed by GitHub
parent 2b7a977cc8
commit 72b89bc88c
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23

View File

@ -0,0 +1,179 @@
package Searches;
import java.util.Collections;
import java.util.ArrayList;
import java.util.Comparator;
import java.util.Random;
/**
* Monte Carlo Tree Search (MCTS) is a heuristic search algorithm
* used in decition taking problems especially games.
*
* See more: https://en.wikipedia.org/wiki/Monte_Carlo_tree_search,
* https://www.baeldung.com/java-monte-carlo-tree-search
*/
public class MonteCarloTreeSearch {
public class Node {
Node parent;
ArrayList <Node> childNodes;
boolean isPlayersTurn; // True if it is the player's turn.
boolean playerWon; // True if the player won; false if the opponent won.
int score;
int visitCount;
public Node() {}
public Node(Node parent, boolean isPlayersTurn) {
this.parent = parent;
childNodes = new ArrayList<>();
this.isPlayersTurn = isPlayersTurn;
playerWon = false;
score = 0;
visitCount = 0;
}
}
static final int WIN_SCORE = 10;
static final int TIME_LIMIT = 500; // Time the algorithm will be running for (in milliseconds).
public static void main(String[] args) {
MonteCarloTreeSearch mcts = new MonteCarloTreeSearch();
mcts.monteCarloTreeSearch(mcts.new Node(null, true));
}
/**
* Explores a game tree using Monte Carlo Tree Search (MCTS)
* and returns the most promising node.
*
* @param rootNode Root node of the game tree.
* @return The most promising child of the root node.
*/
public Node monteCarloTreeSearch(Node rootNode) {
Node winnerNode;
double timeLimit;
// Expand the root node.
addChildNodes(rootNode, 10);
timeLimit = System.currentTimeMillis() + TIME_LIMIT;
// Explore the tree until the time limit is reached.
while (System.currentTimeMillis() < timeLimit) {
Node promisingNode;
// Get a promising node using UCT.
promisingNode = getPromisingNode(rootNode);
// Expand the promising node.
if (promisingNode.childNodes.size() == 0) {
addChildNodes(promisingNode, 10);
}
simulateRandomPlay(promisingNode);
}
winnerNode = getWinnerNode(rootNode);
printScores(rootNode);
System.out.format("\nThe optimal node is: %02d\n", rootNode.childNodes.indexOf(winnerNode) + 1);
return winnerNode;
}
public void addChildNodes(Node node, int childCount) {
for (int i = 0; i < childCount; i++) {
node.childNodes.add(new Node(node, !node.isPlayersTurn));
}
}
/**
* Uses UCT to find a promising child node to be explored.
*
* UCT: Upper Confidence bounds applied to Trees.
*
* @param rootNode Root node of the tree.
* @return The most promising node according to UCT.
*/
public Node getPromisingNode(Node rootNode) {
Node promisingNode = rootNode;
// Iterate until a node that hasn't been expanded is found.
while (promisingNode.childNodes.size() != 0) {
double uctIndex = Double.MIN_VALUE;
int nodeIndex = 0;
// Iterate through child nodes and pick the most promising one
// using UCT (Upper Confidence bounds applied to Trees).
for (int i = 0; i < promisingNode.childNodes.size(); i++) {
Node childNode = promisingNode.childNodes.get(i);
double uctTemp;
// If child node has never been visited
// it will have the highest uct value.
if (childNode.visitCount == 0) {
nodeIndex = i;
break;
}
uctTemp = ((double) childNode.score / childNode.visitCount)
+ 1.41 * Math.sqrt(Math.log(promisingNode.visitCount) / (double) childNode.visitCount);
if (uctTemp > uctIndex) {
uctIndex = uctTemp;
nodeIndex = i;
}
}
promisingNode = promisingNode.childNodes.get(nodeIndex);
}
return promisingNode;
}
/**
* Simulates a random play from a nodes current state
* and back propagates the result.
*
* @param promisingNode Node that will be simulated.
*/
public void simulateRandomPlay(Node promisingNode) {
Random rand = new Random();
Node tempNode = promisingNode;
boolean isPlayerWinner;
// The following line randomly determines whether the simulated play is a win or loss.
// To use the MCTS algorithm correctly this should be a simulation of the nodes' current
// state of the game until it finishes (if possible) and use an evaluation function to
// determine how good or bad the play was.
// e.g. Play tic tac toe choosing random squares until the game ends.
promisingNode.playerWon = (rand.nextInt(6) == 0);
isPlayerWinner = promisingNode.playerWon;
// Back propagation of the random play.
while (tempNode != null) {
tempNode.visitCount++;
// Add wining scores to bouth player and opponent depending on the turn.
if ((tempNode.isPlayersTurn && isPlayerWinner) ||
(!tempNode.isPlayersTurn && !isPlayerWinner)) {
tempNode.score += WIN_SCORE;
}
tempNode = tempNode.parent;
}
}
public Node getWinnerNode(Node rootNode) {
return Collections.max(rootNode.childNodes, Comparator.comparing(c -> c.score));
}
public void printScores(Node rootNode) {
System.out.println("N.\tScore\t\tVisits");
for (int i = 0; i < rootNode.childNodes.size(); i++) {
System.out.println(String.format("%02d\t%d\t\t%d", i + 1,
rootNode.childNodes.get(i).score, rootNode.childNodes.get(i).visitCount));
}
}
}