wfolsom/cs6601p1
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fixed Minimax search to use the new GameTreeNode, MinimaxProperty classes.

Browse Source 
The previous implementation was overly complicated and may have been buggy except when searching only 2 plies ahead.
This commit is contained in:
cs6601
2012-08-30 10:51:04 -04:00
parent 2e40440838
commit 4a1c64843d
12 changed files with 249 additions and 153 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
src/net/woodyfolsom/msproj/policy/MonteCarloUCT.java
View File

@@ -4,6 +4,8 @@ import java.util.ArrayList;
import java.util.List;
import net.woodyfolsom.msproj.Action;
import net.woodyfolsom.msproj.tree.GameTreeNode;
import net.woodyfolsom.msproj.tree.MonteCarloProperties;
public class MonteCarloUCT extends MonteCarlo {
@@ -12,17 +14,20 @@ public class MonteCarloUCT extends MonteCarlo {
}
@Override
public List<GameTreeNode> descend(GameTreeNode node) {
double bestScore = (double) node.getNumWins() / node.getNumVisits();
GameTreeNode bestNode = node;
public List<GameTreeNode<MonteCarloProperties>> descend(GameTreeNode<MonteCarloProperties> node) {
double bestScore = Double.NEGATIVE_INFINITY;
GameTreeNode<MonteCarloProperties> bestNode = node;
//This appears slightly redundant with getBestAction() but it is not -
//descend() may pick the current node rather than a child to expand (if a child has a good score but high/low uncertainty)
//but getBestAction specifically asks for the optimum action to take from the current node,
//even if it results in a worse next state.
for (Action action : node.getActions()) {
GameTreeNode childNode = node.getChild(action);
double childScore = (double) childNode.getNumWins() / childNode.getNumVisits();
GameTreeNode<MonteCarloProperties> childNode = node.getChild(action);
MonteCarloProperties properties = childNode.getProperties();
double childScore = (double) properties.getWins() / properties.getVisits();
if (childScore >= bestScore) {
bestScore = childScore;
bestNode = childNode;
@@ -30,7 +35,7 @@ public class MonteCarloUCT extends MonteCarlo {
}
if (bestNode == node) {
List<GameTreeNode> bestNodeList = new ArrayList<GameTreeNode>();
List<GameTreeNode<MonteCarloProperties>> bestNodeList = new ArrayList<GameTreeNode<MonteCarloProperties>>();
bestNodeList.add(bestNode);
return bestNodeList;
} else {
@@ -39,13 +44,16 @@ public class MonteCarloUCT extends MonteCarlo {
}
@Override
public Action getBestAction(GameTreeNode node) {
public Action getBestAction(GameTreeNode<MonteCarloProperties> node) {
Action bestAction = Action.NONE;
double bestScore = Double.NEGATIVE_INFINITY;
for (Action action : node.getActions()) {
GameTreeNode childNode = node.getChild(action);
double childScore = (double) childNode.getNumWins() / childNode.getNumVisits();
GameTreeNode<MonteCarloProperties> childNode = node.getChild(action);
MonteCarloProperties properties = childNode.getProperties();
double childScore = (double) properties.getWins() / properties.getVisits();
if (childScore >= bestScore) {
bestScore = childScore;
bestAction = action;
@@ -56,19 +64,19 @@ public class MonteCarloUCT extends MonteCarlo {
}
@Override
public List<GameTreeNode> grow(GameTreeNode node) {
public List<GameTreeNode<MonteCarloProperties>> grow(GameTreeNode<MonteCarloProperties> node) {
// TODO Auto-generated method stub
return null;
}
@Override
public int rollout(GameTreeNode node) {
public int rollout(GameTreeNode<MonteCarloProperties> node) {
// TODO Auto-generated method stub
return 0;
}
@Override
public void update(GameTreeNode node, int reward) {
public void update(GameTreeNode<MonteCarloProperties> node, int reward) {
// TODO Auto-generated method stub
}