diff --git a/mcts.py b/mcts.py
new file mode 100644
index 0000000..946a42e
--- /dev/null
+++ b/mcts.py
@@ -0,0 +1,65 @@
+from node import MCTSNode
+from board import GomokuBoard
+import numpy as np
+import copy
+import torch
+
+class MCTS:
+    def __init__(self, policy_value_fn, c_puct=1.0, n_playout=100):
+        self.root = MCTSNode(None, 1.0)
+        self.policy_value_fn = policy_value_fn
+        self.c_puct = c_puct
+        self.n_playout = n_playout
+
+    def playout(self, state):
+        node = self.root
+        board = copy.deepcopy(state)  # 这里改成用传入的状态深拷贝
+
+        # 选择阶段
+        while not node.is_leaf():
+            action, node = node.select(self.c_puct)
+            board.execute_move(action)
+
+        # 评估阶段
+        board_state = board.get_board_state()
+        if board_state == 1:  # 黑赢
+            value = 1.0
+        elif board_state == -1:  # 白赢
+            value = -1.0
+        elif board_state == 2:  # 平局
+            value = 0.0
+        else:
+            
+            input_tensor = board.get_board_data()  # numpy array
+            input_tensor = torch.from_numpy(input_tensor).float()  # 转成float tensor
+            if torch.cuda.is_available():
+                input_tensor = input_tensor.cuda()
+            input_tensor = input_tensor  # 加 batch 维度 (1, C, H, W)
+
+            act_probs, value = self.policy_value_fn(input_tensor)
+
+            legal_moves = board.get_available_move()
+            legal_indices = [i for i, flag in enumerate(legal_moves) if flag == 1]
+            probs = [(i, act_probs[0][i]) for i in legal_indices]
+            node.expand(probs)
+
+        # 回传阶段
+        node.update_recursive(-value)
+
+
+    def get_move_probs(self, state, temp=1e-3):
+        for _ in range(self.n_playout):
+            self.playout(state)
+
+        act_visits = [(act, node.N) for act, node in self.root.children.items()]
+        acts, visits = zip(*act_visits)
+        act_probs = np.power(visits, 1.0 / temp)
+        act_probs /= np.sum(act_probs)
+        return acts, act_probs
+
+    def update_with_move(self, last_move):
+        if last_move in self.root.children:
+            self.root = self.root.children[last_move]
+            self.root.parent = None
+        else:
+            self.root = MCTSNode(None, 1.0)