[Example] Add MCTS example

ghstack-source-id: 4cf2a16
Pull Request resolved: pytorch#2796

Author: kurtamohler

examples/trees/mcts.py

@@ -0,0 +1,232 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import time
+
+import torch
+import torchrl
+import torchrl.envs
+from tensordict import TensorDict
+
+from torchrl.data.map import (MCTSForest, Tree)
+from tensordict import TensorDictBase
+from torchrl.envs import EnvBase
+
+start_time = time.time()
+
+pgn_or_fen = "fen"
+mask_actions = True
+
+env = torchrl.envs.ChessEnv(
+    include_pgn=False,
+    include_fen=True,
+    include_hash=True,
+    include_hash_inv=True,
+    include_san=True,
+    stateful=True,
+    mask_actions=mask_actions,
+)
+
+class TransformReward:
+    def __init__(self):
+        self.first_turn = None
+
+    def __call__(self, td):
+        if self.first_turn is None and "turn" in td:
+            self.first_turn = td["turn"]
+            print(f'first turn: {self.first_turn}')
+
+        if "reward" not in td:
+            return td
+        reward = td["reward"]
+        if reward == 0.5:
+            reward = 0
+        #elif reward == 1 and td["turn"] == env.lib.WHITE:
+        elif reward == 1 and td["turn"] == self.first_turn:
+            reward = -reward
+
+        td["reward"] = reward
+        return td
+
+    def reset(self, td):
+        self.first_turn = None
+
+
+# ChessEnv sets the reward to 0.5 for a draw and 1 for a win for either player.
+# Need to transform the reward to be:
+#   white win = 1
+#   draw = 0
+#   black win = -1
+env = env.append_transform(TransformReward())
+
+forest = torchrl.data.MCTSForest()
+forest.reward_keys = env.reward_keys
+forest.done_keys = env.done_keys
+forest.action_keys = env.action_keys
+
+if mask_actions:
+    forest.observation_keys = [f"{pgn_or_fen}_hash", "turn", "action_mask"]
+else:
+    forest.observation_keys = [f"{pgn_or_fen}_hash", "turn"]
+
+C = 2.0**0.5
+
+
+def traversal_priority_UCB1(tree):
+    subtree = tree.subtree
+    visits = subtree.visits
+    reward_sum = subtree.wins
+
+    # If it's black's turn, flip the reward, since black wants to
+    # optimize for the lowest reward, not highest.
+    if not subtree.rollout[0, 0]["turn"]:
+        reward_sum = -reward_sum
+
+    parent_visits = tree.visits
+    reward_sum = reward_sum.squeeze(-1)
+    priority = (reward_sum + C * torch.sqrt(torch.log(parent_visits))) / visits
+    priority[visits == 0] = float("inf")
+    return priority
+
+
+def _traverse_MCTS_one_step(forest, tree, env, max_rollout_steps):
+    done = False
+    trees_visited = [tree]
+
+    while not done:
+        if tree.subtree is None:
+            td_tree = tree.rollout[-1]["next"].clone()
+
+            if (tree.visits > 0 or tree.parent is None) and not td_tree["done"]:
+                actions = env.all_actions(td_tree)
+                subtrees = []
+
+                for action in actions:
+                    td = env.step(env.reset(td_tree).update(action))
+                    new_node = torchrl.data.Tree(
+                        rollout=td.unsqueeze(0),
+                        node_data=td["next"].select(*forest.node_map.in_keys),
+                        count=torch.tensor(0),
+                        wins=torch.zeros_like(td["next", env.reward_key]),
+                    )
+                    subtrees.append(new_node)
+
+                # NOTE: This whole script runs about 2x faster with lazy stack
+                # versus eager stack.
+                tree.subtree = TensorDict.lazy_stack(subtrees)
+                chosen_idx = torch.randint(0, len(subtrees), ()).item()
+                rollout_state = subtrees[chosen_idx].rollout[-1]["next"]
+
+            else:
+                rollout_state = td_tree
+
+            if rollout_state["done"]:
+                rollout_reward = rollout_state[env.reward_key]
+            else:
+                rollout = env.rollout(
+                    max_steps=max_rollout_steps,
+                    tensordict=rollout_state,
+                )
+                rollout_reward = rollout[-1]["next", env.reward_key]
+            done = True
+
+        else:
+            priorities = traversal_priority_UCB1(tree)
+            chosen_idx = torch.argmax(priorities).item()
+            tree = tree.subtree[chosen_idx]
+            trees_visited.append(tree)
+
+    for tree in trees_visited:
+        tree.visits += 1
+        tree.wins += rollout_reward
+
+
+def MCTS(
+    forest: MCTSForest,
+    root: TensorDictBase,
+    env: EnvBase,
+    num_steps: int,
+    max_rollout_steps: int | None = None
+) -> Tree:
+    """Performs Monte-Carlo tree search in an environment.
+
+    Args:
+        forest (MCTSForest): Forest of the tree to update. If the tree does not
+            exist yet, it is added.
+        root (TensorDict): The root step of the tree to update.
+        env (EnvBase): Environment to performs actions in.
+        num_steps (int): Number of iterations to traverse.
+        max_rollout_steps (int): Maximum number of steps for each rollout.
+    """
+    if root not in forest:
+        for action in env.all_actions(root):
+            td = env.step(env.reset(root.clone()).update(action))
+            forest.extend(td.unsqueeze(0))
+
+    tree = forest.get_tree(root)
+    tree.wins = torch.zeros_like(td["next", env.reward_key])
+    for subtree in tree.subtree:
+        subtree.wins = torch.zeros_like(td["next", env.reward_key])
+
+    for _ in range(num_steps):
+        _traverse_MCTS_one_step(forest, tree, env, max_rollout_steps)
+
+    return tree
+
+
+def tree_format_fn(tree):
+    td = tree.rollout[-1]["next"]
+    return [
+        td["san"],
+        td[pgn_or_fen].split("\n")[-1],
+        tree.wins,
+        tree.visits,
+    ]
+
+
+def get_best_move(fen, mcts_steps, rollout_steps):
+    root = env.reset(TensorDict({"fen": fen}))
+    tree = MCTS(forest, root, env, mcts_steps, rollout_steps)
+    moves = []
+
+    for subtree in tree.subtree:
+        san = subtree.rollout[0]["next", "san"]
+        reward_sum = subtree.wins
+        visits = subtree.visits
+        value_avg = (reward_sum / visits).item()
+
+        if not subtree.rollout[0]["turn"]:
+            value_avg = -value_avg
+
+        moves.append((value_avg, san))
+
+    moves = sorted(moves, key=lambda x: -x[0])
+
+    #print(tree.to_string(tree_format_fn))
+
+    print("------------------")
+    for value_avg, san in moves:
+        print(f" {value_avg:0.02f} {san}")
+    print("------------------")
+
+    return moves[0][1]
+
+
+# White has M1, best move Rd8#. Any other moves lose to M2 or M1.
+fen0 = "7k/6pp/7p/7K/8/8/6q1/3R4 w - - 0 1"
+assert get_best_move(fen0, 100, 10) == "Rd8#"
+
+# Black has M1, best move Qg6#. Other moves give rough equality or worse.
+fen1 = "6qk/2R4p/7K/8/8/8/8/4R3 b - - 1 1"
+assert get_best_move(fen1, 100, 10) == "Qg6#"
+
+# White has M2, best move Rxg8+. Any other move loses.
+fen2 = "2R3qk/5p1p/7K/8/8/8/5r2/2R5 w - - 0 1"
+assert get_best_move(fen2, 1000, 10) == "Rxg8+"
+
+end_time = time.time()
+total_time = end_time - start_time
+
+print(f"Took {total_time} s")

torchrl/data/map/tree.py

@@ -1363,6 +1363,11 @@ def valid_paths(cls, tree: Tree):
     def __len__(self):
         return len(self.data_map)
 
+    def __contains__(self, root: TensorDictBase):
+        if self.node_map is None:
+            return False
+        return root.select(*self.node_map.in_keys) in self.node_map
+
     def to_string(self, td_root, node_format_fn=lambda tree: tree.node_data.to_dict()):
         """Generates a string representation of a tree in the forest.
 

torchrl/envs/custom/chess.py

@@ -222,12 +222,15 @@ def lib(cls):
         return chess
 
     _san_moves = []
+    _san_move_to_index_map = {}
 
     @_classproperty
     def san_moves(cls):
         if not cls._san_moves:
             with open(pathlib.Path(__file__).parent / "san_moves.txt", "r+") as f:
                 cls._san_moves.extend(f.read().split("\n"))
+            for idx, san_move in enumerate(cls._san_moves):
+                cls._san_move_to_index_map[san_move] = idx
         return cls._san_moves
 
     def _legal_moves_to_index(
@@ -255,7 +258,7 @@ def _legal_moves_to_index(
             board = self.board
 
         indices = torch.tensor(
-            [self._san_moves.index(board.san(m)) for m in board.legal_moves],
+            [self._san_move_to_index_map[board.san(m)] for m in board.legal_moves],
             dtype=torch.int64,
         )
         mask = None
@@ -409,7 +412,9 @@ def _reset(self, tensordict=None):
             if move is None:
                 dest.set("san", "<start>")
             else:
-                dest.set("san", self.board.san(move))
+                prev_board = self.board.copy()
+                prev_board.pop()
+                dest.set("san", prev_board.san(move))
         if self.include_fen:
             dest.set("fen", fen)
         if self.include_pgn: