Bonus Tutorial: Planning with Monte Carlo Tree Search

Bonus Tutorial: Planning with Monte Carlo Tree Search#

Week 3, Day 5: Reinforcement Learning for Games & DL Thinking 3

By Neuromatch Academy

Content creators: Mandana Samiei, Raymond Chua, Kushaan Gupta, Tim Lilicrap, Blake Richards

Content reviewers: Arush Tagade, Lily Cheng, Melvin Selim Atay, Kelson Shilling-Scrivo

Content editors: Melvin Selim Atay, Spiros Chavlis, Gunnar Blohm

Production editors: Namrata Bafna, Gagana B, Spiros Chavlis

Tutorial Objectives#

In this tutorial, you will learn about Monte Carlo Tree Search (MCTS) and compare its performance to policy-based, value-based players, and Monte Carlo planners.

These are the slides for the videos in the tutorial. If you want to locally download the slides, click here.

Setup#

Install dependencies#

Install and import feedback gadget#

# Imports
import os
import math
import random
import time
import torch
import random
import logging
import coloredlogs

import numpy as np
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F

from tqdm.notebook import tqdm

log = logging.getLogger(__name__)
coloredlogs.install(level='INFO')  # Change this to DEBUG to see more info.

Set random seed#

Executing set_seed(seed=seed) you are setting the seed

Set device (GPU or CPU). Execute `set_device()`#

SEED = 2023
set_seed(seed=SEED)
DEVICE = set_device()

Random seed 2023 has been set.
WARNING: For this notebook to perform best, if possible, in the menu under `Runtime` -> `Change runtime type.`  select `GPU` 

Download the modules#

Add the nma_rl_games in the path and import the modules.

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
Cell In[8], line 38
     36 from othello.OthelloLogic import Board
     37 # from othello.OthelloGame import OthelloGame
---> 38 from othello.pytorch.NNet import NNetWrapper as NNet

File ~/work/course-content-dl/course-content-dl/tutorials/W3D5_ReinforcementLearningForGamesAndDlThinking3/student/nma_rl_games/alpha-zero/othello/pytorch/NNet.py:15
     12 import torch
     13 import torch.optim as optim
---> 15 from .OthelloNNet import OthelloNNet as onnet
     17 args = dotdict({
     18     'lr': 0.001,
     19     'dropout': 0.3,
   (...)
     23     'num_channels': 512,
     24 })
     27 class NNetWrapper(NeuralNet):

File ~/work/course-content-dl/course-content-dl/tutorials/W3D5_ReinforcementLearningForGamesAndDlThinking3/student/nma_rl_games/alpha-zero/othello/pytorch/OthelloNNet.py:10
      8 import torch.nn.functional as F
      9 import torch.optim as optim
---> 10 from torchvision import datasets, transforms
     11 from torch.autograd import Variable
     13 class OthelloNNet(nn.Module):

File /opt/hostedtoolcache/Python/3.9.20/x64/lib/python3.9/site-packages/torchvision/__init__.py:6
      3 from modulefinder import Module
      5 import torch
----> 6 from torchvision import _meta_registrations, datasets, io, models, ops, transforms, utils
      8 from .extension import _HAS_OPS
     10 try:

File /opt/hostedtoolcache/Python/3.9.20/x64/lib/python3.9/site-packages/torchvision/_meta_registrations.py:164
    153     torch._check(
    154         grad.dtype == rois.dtype,
    155         lambda: (
   (...)
    158         ),
    159     )
    160     return grad.new_empty((batch_size, channels, height, width))
    163 @torch._custom_ops.impl_abstract("torchvision::nms")
--> 164 def meta_nms(dets, scores, iou_threshold):
    165     torch._check(dets.dim() == 2, lambda: f"boxes should be a 2d tensor, got {dets.dim()}D")
    166     torch._check(dets.size(1) == 4, lambda: f"boxes should have 4 elements in dimension 1, got {dets.size(1)}")

File /opt/hostedtoolcache/Python/3.9.20/x64/lib/python3.9/site-packages/torch/library.py:795, in register_fake.<locals>.register(func)
    793 else:
    794     use_lib = lib
--> 795 use_lib._register_fake(op_name, func, _stacklevel=stacklevel + 1)
    796 return func

File /opt/hostedtoolcache/Python/3.9.20/x64/lib/python3.9/site-packages/torch/library.py:184, in Library._register_fake(self, op_name, fn, _stacklevel)
    181 else:
    182     func_to_register = fn
--> 184 handle = entry.fake_impl.register(func_to_register, source)
    185 self._registration_handles.append(handle)

File /opt/hostedtoolcache/Python/3.9.20/x64/lib/python3.9/site-packages/torch/_library/fake_impl.py:31, in FakeImplHolder.register(self, func, source)
     25 if self.kernel is not None:
     26     raise RuntimeError(
     27         f"register_fake(...): the operator {self.qualname} "
     28         f"already has an fake impl registered at "
     29         f"{self.kernel.source}."
     30     )
---> 31 if torch._C._dispatch_has_kernel_for_dispatch_key(self.qualname, "Meta"):
     32     raise RuntimeError(
     33         f"register_fake(...): the operator {self.qualname} "
     34         f"already has an DispatchKey::Meta implementation via a "
   (...)
     37         f"register_fake."
     38     )
     40 if torch._C._dispatch_has_kernel_for_dispatch_key(
     41     self.qualname, "CompositeImplicitAutograd"
     42 ):

RuntimeError: operator torchvision::nms does not exist

Helper functions from previous tutorials#

Show code cell source Hide code cell source

# @title Helper functions from previous tutorials

def loadTrainExamples(folder, filename):
  """
  Helper function to load training examples

  Args:
    folder: string
      Path specifying training examples
    filename: string
      File name of training examples

  Returns:
    trainExamplesHistory: list
      Returns examples based on the model were already collected (loaded)
  """
  trainExamplesHistory = []
  modelFile = os.path.join(folder, filename)
  examplesFile = modelFile + ".examples"
  if not os.path.isfile(examplesFile):
    print(f'File "{examplesFile}" with trainExamples not found!')
    r = input("Continue? [y|n]")
    if r != "y":
      sys.exit()
  else:
    print("File with train examples found. Loading it...")
    with open(examplesFile, "rb") as f:
      trainExamplesHistory = Unpickler(f).load()
    print('Loading done!')
    return trainExamplesHistory


def save_model_checkpoint(folder, filename, nnet):
  filepath = os.path.join(folder, filename)

  if not os.path.exists(folder):
    print("Checkpoint Directory does not exist! Making directory {}".format(folder))
    os.mkdir(folder)
  else:
    print("Checkpoint Directory exists!")

  torch.save({'state_dict': nnet.state_dict()}, filepath)
  print("Model saved!")

def load_model_checkpoint(folder, filename, nnet, device):
  filepath = os.path.join(folder, filename)

  if not os.path.exists(filepath):
    raise FileNotFoundError("No model in path {}".format(filepath))

  checkpoint = torch.load(filepath, map_location=device)
  nnet.load_state_dict(checkpoint['state_dict'])


class OthelloGame(Game):
  """
  Othello game board
  """
  square_content = {
      -1: "X",
      +0: "-",
      +1: "O"
      }

  @staticmethod
  def getSquarePiece(piece):
    return OthelloGame.square_content[piece]

  def __init__(self, n):
    self.n = n

  def getInitBoard(self):
    b = Board(self.n)
    return np.array(b.pieces)

  def getBoardSize(self):
    return (self.n, self.n)

  def getActionSize(self):
    # Return number of actions, n is the board size and +1 is for no-op action
    return self.n * self.n + 1

  def getCanonicalForm(self, board, player):
    # Return state if player==1, else return -state if player==-1
    return player * board

  def stringRepresentation(self, board):
    return board.tobytes()

  def stringRepresentationReadable(self, board):
    board_s = "".join(self.square_content[square] for row in board for square in row)
    return board_s

  def getScore(self, board, player):
    b = Board(self.n)
    b.pieces = np.copy(board)
    return b.countDiff(player)

  @staticmethod
  def display(board):
    n = board.shape[0]
    print("   ", end="")
    for y in range(n):
      print(y, end=" ")
    print("")
    print("-----------------------")
    for y in range(n):
      print(y, "|", end="")    # Print the row
      for x in range(n):
        piece = board[y][x]    # Get the piece to print
        print(OthelloGame.square_content[piece], end=" ")
      print("|")
    print("-----------------------")

  @staticmethod
  def displayValidMoves(moves):
    A=np.reshape(moves[0:-1], board.shape)
    n = board.shape[0]
    print("  ")
    print("possible moves")
    print("   ", end="")
    for y in range(n):
      print(y, end=" ")
    print("")
    print("-----------------------")
    for y in range(n):
      print(y, "|", end="")    # Print the row
      for x in range(n):
        piece = A[y][x]    # Get the piece to print
        print(OthelloGame.square_content[piece], end=" ")
      print("|")
    print("-----------------------")

  def getNextState(self, board, player, action):
    """
    Make valid move. If player takes action on board, return next (board,player)
    and action must be a valid move

    Args:
      board: np.ndarray
        Board of size n x n [6x6 in this case]
      player: Integer
        ID of current player
      action: np.ndarray
        Space of actions

    Returns:
      (board, player): tuple
        Next state representation
    """
    if action == self.n*self.n:
      return (board, -player)
    b = Board(self.n)
    b.pieces = np.copy(board)
    move = (int(action/self.n), action%self.n)
    b.execute_move(move, player)
    return (b.pieces, -player)

  def getValidMoves(self, board, player):
    """
    Get all valid moves for player

    Args:
      board: np.ndarray
        Board of size n x n [6x6 in this case]
      player: Integer
        ID of current player
      action: np.ndarray
        Space of action

    Returns:
      valids: np.ndarray
        Valid moves for player
    """
    valids = [0]*self.getActionSize()
    b = Board(self.n)
    b.pieces = np.copy(board)
    legalMoves =  b.get_legal_moves(player)
    if len(legalMoves)==0:
      valids[-1]=1
      return np.array(valids)
    for x, y in legalMoves:
      valids[self.n*x+y]=1
    return np.array(valids)

  def getGameEnded(self, board, player):
    """
    Check if game ended

    Args:
      board: np.ndarray
        Board of size n x n [6x6 in this case]
      player: Integer
        ID of current player

    Returns:
      0 if not ended, 1 if player 1 won, -1 if player 1 lost
    """
    b = Board(self.n)
    b.pieces = np.copy(board)
    if b.has_legal_moves(player):
      return 0
    if b.has_legal_moves(-player):
      return 0
    if b.countDiff(player) > 0:
      return 1
    return -1

  def getSymmetries(self, board, pi):
    """
    Get mirror/rotational configurations of board

    Args:
      board: np.ndarray
        Board of size n x n [6x6 in this case]
      pi: np.ndarray
        Dimension of board

    Returns:
      l: list
        90 degree of board, 90 degree of pi_board
    """
    assert(len(pi) == self.n**2+1)  # 1 for pass
    pi_board = np.reshape(pi[:-1], (self.n, self.n))
    l = []

    for i in range(1, 5):
      for j in [True, False]:
        newB = np.rot90(board, i)
        newPi = np.rot90(pi_board, i)
        if j:
          newB = np.fliplr(newB)
          newPi = np.fliplr(newPi)
        l += [(newB, list(newPi.ravel()) + [pi[-1]])]
    return l


class RandomPlayer():

  def __init__(self, game):
    self.game = game

  def play(self, board):
    """
    Simulates game play

    Args:
      board: np.ndarray
        Board of size n x n [6x6 in this case]

    Returns:
      a: int
        Randomly chosen move
    """

    # Compute the valid moves using getValidMoves()
    valids = self.game.getValidMoves(board, 1)

    # Compute the probability of each move being played (random player means this should
    # be uniform for valid moves, 0 for others)
    prob = valids/valids.sum()

    # Pick an action based on the probabilities (hint: np.choice is useful)
    a = np.random.choice(self.game.getActionSize(), p=prob)

    return a


class OthelloNNet(nn.Module):

  def __init__(self, game, args):
    """
    Initialise game parameters

    Args:
      game: OthelloGame instance
        Instance of the OthelloGame class above;
      args: dictionary
        Instantiates number of iterations and episodes, controls temperature threshold, queue length,
        arena, checkpointing, and neural network parameters:
        learning-rate: 0.001, dropout: 0.3, epochs: 10, batch_size: 64,
        num_channels: 512
    """
    self.board_x, self.board_y = game.getBoardSize()
    self.action_size = game.getActionSize()
    self.args = args

    super(OthelloNNet, self).__init__()
    self.conv1 = nn.Conv2d(in_channels=1, out_channels=args.num_channels,
                           kernel_size=3, stride=1, padding=1)
    self.conv2 = nn.Conv2d(in_channels=args.num_channels,
                           out_channels=args.num_channels, kernel_size=3,
                           stride=1, padding=1)
    self.conv3 = nn.Conv2d(in_channels=args.num_channels,
                           out_channels=args.num_channels, kernel_size=3,
                           stride=1)
    self.conv4 = nn.Conv2d(in_channels=args.num_channels,
                           out_channels=args.num_channels, kernel_size=3,
                           stride=1)

    self.bn1 = nn.BatchNorm2d(num_features=args.num_channels)
    self.bn2 = nn.BatchNorm2d(num_features=args.num_channels)
    self.bn3 = nn.BatchNorm2d(num_features=args.num_channels)
    self.bn4 = nn.BatchNorm2d(num_features=args.num_channels)

    self.fc1 = nn.Linear(in_features=args.num_channels * (self.board_x - 4) * (self.board_y - 4),
                         out_features=1024)
    self.fc_bn1 = nn.BatchNorm1d(num_features=1024)

    self.fc2 = nn.Linear(in_features=1024, out_features=512)
    self.fc_bn2 = nn.BatchNorm1d(num_features=512)

    self.fc3 = nn.Linear(in_features=512, out_features=self.action_size)

    self.fc4 = nn.Linear(in_features=512, out_features=1)

  def forward(self, s):
    """
    Controls forward pass of OthelloNNet

    Args:
      s: np.ndarray
        Array of size (batch_size x board_x x board_y)

    Returns:
      prob, v: tuple of torch.Tensor
        Probability distribution over actions at the current state and the value
        of the current state.
    """
    s = s.view(-1, 1, self.board_x, self.board_y)                # batch_size x 1 x board_x x board_y
    s = F.relu(self.bn1(self.conv1(s)))                          # batch_size x num_channels x board_x x board_y
    s = F.relu(self.bn2(self.conv2(s)))                          # batch_size x num_channels x board_x x board_y
    s = F.relu(self.bn3(self.conv3(s)))                          # batch_size x num_channels x (board_x-2) x (board_y-2)
    s = F.relu(self.bn4(self.conv4(s)))                          # batch_size x num_channels x (board_x-4) x (board_y-4)
    s = s.view(-1, self.args.num_channels * (self.board_x - 4) * (self.board_y - 4))

    s = F.dropout(F.relu(self.fc_bn1(self.fc1(s))), p=self.args.dropout, training=self.training)  # batch_size x 1024
    s = F.dropout(F.relu(self.fc_bn2(self.fc2(s))), p=self.args.dropout, training=self.training)  # batch_size x 512

    pi = self.fc3(s)  # batch_size x action_size
    v = self.fc4(s)   # batch_size x 1

    return F.log_softmax(pi, dim=1), torch.tanh(v)


class ValueNetwork(NeuralNet):

  def __init__(self, game):
    """
    Args:
      game: OthelloGame
        Instance of the OthelloGame class above
    """
    self.nnet = OthelloNNet(game, args)
    self.board_x, self.board_y = game.getBoardSize()
    self.action_size = game.getActionSize()
    self.nnet.to(args.device)

  def train(self, games):
    """
    Args:
      games: list
        List of examples with each example is of form (board, pi, v)
    """
    optimizer = optim.Adam(self.nnet.parameters())
    for examples in games:
      for epoch in range(args.epochs):
        print('EPOCH ::: ' + str(epoch + 1))
        self.nnet.train()
        v_losses = []   # To store the losses per epoch
        batch_count = int(len(examples) / args.batch_size)  # len(examples)=200, batch-size=64, batch_count=3
        t = tqdm(range(batch_count), desc='Training Value Network')
        for _ in t:
          sample_ids = np.random.randint(len(examples), size=args.batch_size)  # Read the ground truth information from MCTS simulation using the loaded examples
          boards, pis, vs = list(zip(*[examples[i] for i in sample_ids]))  # Length of boards, pis, vis = 64
          boards = torch.FloatTensor(np.array(boards).astype(np.float64))
          target_vs = torch.FloatTensor(np.array(vs).astype(np.float64))

          # Predict
          # To run on GPU if available
          boards, target_vs = boards.contiguous().to(args.device), target_vs.contiguous().to(args.device)

          # Compute output
          _, out_v = self.nnet(boards)
          l_v = self.loss_v(target_vs, out_v)  # Total loss

          # Record loss
          v_losses.append(l_v.item())
          t.set_postfix(Loss_v=l_v.item())

          # Compute gradient and do SGD step
          optimizer.zero_grad()
          l_v.backward()
          optimizer.step()

  def predict(self, board):
    """
    Args:
      board: np.ndarray
        Board of size n x n [6x6 in this case]

    Returns:
      v: OthelloNet instance
        Data of the OthelloNet class instance above;
    """
    # Timing
    start = time.time()

    # Preparing input
    board = torch.FloatTensor(board.astype(np.float64))
    board = board.contiguous().to(args.device)
    board = board.view(1, self.board_x, self.board_y)
    self.nnet.eval()
    with torch.no_grad():
        _, v = self.nnet(board)
    return v.data.cpu().numpy()[0]

  def loss_v(self, targets, outputs):
    """
    Args:
      targets: np.ndarray
        Ground Truth variables corresponding to input
      outputs: np.ndarray
        Predictions of Network

    Returns:
      MSE Loss averaged across the whole dataset
    """
    # Mean squared error (MSE)
    return torch.sum((targets - outputs.view(-1)) ** 2) / targets.size()[0]

  def save_checkpoint(self, folder='checkpoint', filename='checkpoint.pth.tar'):
    save_model_checkpoint(folder, filename, self.nnet)

  def load_checkpoint(self, folder='checkpoint', filename='checkpoint.pth.tar'):
    load_model_checkpoint(folder, filename, self.nnet, args.device)

class ValueBasedPlayer():

  def __init__(self, game, vnet):
    """
    Args:
      game: OthelloGame instance
        Instance of the OthelloGame class
      vnet: Value Network instance
        Instance of the Value Network class
    """
    self.game = game
    self.vnet = vnet

  def play(self, board):
    """
    Args:
      board: np.ndarray
        Board of size n x n [6x6 in this case]

    Returns:
      candidates: List
        Collection of tuples describing action and values of future predicted
        states
    """
    valids = self.game.getValidMoves(board, 1)
    candidates = []
    max_num_actions = 4
    va = np.where(valids)[0]
    va_list = va.tolist()
    random.shuffle(va_list)
    for a in va_list:
      # Return next board state using getNextState() function
      nextBoard, _ = self.game.getNextState(board, 1, a)
      # Predict the value of next state using value network
      value = self.vnet.predict(nextBoard)
      # Add the value and the action as a tuple to the candidate lists, note that you might need to change the sign of the value based on the player
      candidates += [(-value, a)]

      if len(candidates) == max_num_actions:
        break

    # Sort by the values
    candidates.sort()

    # Return action associated with highest value
    return candidates[0][1]


class PolicyNetwork(NeuralNet):

  def __init__(self, game):
    """
    Args:
      game: OthelloGame
        Instance of the OthelloGame class
    """
    self.nnet = OthelloNNet(game, args)
    self.board_x, self.board_y = game.getBoardSize()
    self.action_size = game.getActionSize()
    self.nnet.to(args.device)

  def train(self, games):
    """
    Args:
      games: list
        List of examples where each example is of form (board, pi, v)
    """
    optimizer = optim.Adam(self.nnet.parameters())

    for examples in games:
      for epoch in range(args.epochs):
        print('EPOCH ::: ' + str(epoch + 1))
        self.nnet.train()
        pi_losses = []

        batch_count = int(len(examples) / args.batch_size)

        t = tqdm(range(batch_count), desc='Training Policy Network')
        for _ in t:
          sample_ids = np.random.randint(len(examples), size=args.batch_size)
          boards, pis, _ = list(zip(*[examples[i] for i in sample_ids]))
          boards = torch.FloatTensor(np.array(boards).astype(np.float64))
          target_pis = torch.FloatTensor(np.array(pis))

          # Predict
          boards, target_pis = boards.contiguous().to(args.device), target_pis.contiguous().to(args.device)

          # Compute output
          out_pi, _ = self.nnet(boards)
          l_pi = self.loss_pi(target_pis, out_pi)

          # Record loss
          pi_losses.append(l_pi.item())
          t.set_postfix(Loss_pi=l_pi.item())

          # Compute gradient and do SGD step
          optimizer.zero_grad()
          l_pi.backward()
          optimizer.step()

  def predict(self, board):
    """
    Args:
      board: np.ndarray
        Board of size n x n [6x6 in this case]

    Returns:
      Data from the OthelloNet instance
    """
    # Timing
    start = time.time()

    # Preparing input
    board = torch.FloatTensor(board.astype(np.float64))
    board = board.contiguous().to(args.device)
    board = board.view(1, self.board_x, self.board_y)
    self.nnet.eval()
    with torch.no_grad():
      pi,_ = self.nnet(board)
    return torch.exp(pi).data.cpu().numpy()[0]

  def loss_pi(self, targets, outputs):
    """
    Calculates Negative Log Likelihood(NLL) of Targets

    Args:
      targets: np.ndarray
        Ground Truth variables corresponding to input
      outputs: np.ndarray
        Predictions of Network

    Returns:
      Negative Log Likelihood calculated as: When training a model, we aspire to
      find the minima of a loss function given a set of parameters (in a neural
      network, these are the weights and biases).
      Sum the loss function to all the correct classes. So, whenever the network
      assigns high confidence at the correct class, the NLL is low, but when the
      network assigns low confidence at the correct class, the NLL is high.
    """
    ## For more information, here is a reference that connects the expression to
    # the neg-log-prob: https://gombru.github.io/2018/05/23/cross_entropy_loss/
    return -torch.sum(targets * outputs) / targets.size()[0]

  def save_checkpoint(self, folder='checkpoint', filename='checkpoint.pth.tar'):
    save_model_checkpoint(folder, filename, self.nnet)

  def load_checkpoint(self, folder='checkpoint', filename='checkpoint.pth.tar'):
    load_model_checkpoint(folder, filename, self.nnet, args.device)


class PolicyBasedPlayer():

  def __init__(self, game, pnet, greedy=True):
    """
    Args:
      game: OthelloGame instance
        Instance of the OthelloGame class above;
      pnet: Policy Network instance
        Instance of the Policy Network class above
      greedy: Boolean
        If true, implement greedy approach
        Else, implement random sample policy based player
    """
    self.game = game
    self.pnet = pnet
    self.greedy = greedy

  def play(self, board):
    """
    Args:
      board: np.ndarray
        Board of size n x n [6x6 in this case]

    Returns:
      a: np.ndarray
        If greedy, implement greedy policy player
        Else, implement random sample policy based player
    """
    valids = self.game.getValidMoves(board, 1)
    action_probs = self.pnet.predict(board)
    vap = action_probs*valids  # Masking invalid moves
    sum_vap = np.sum(vap)

    if sum_vap > 0:
      vap /= sum_vap  # Renormalize
    else:
      # If all valid moves were masked we make all valid moves equally probable
      print("All valid moves were masked, doing a workaround.")
      vap = vap + valids
      vap /= np.sum(vap)

    if self.greedy:
      # Greedy policy player
      a = np.where(vap == np.max(vap))[0][0]
    else:
      # Sample-based policy player
      a = np.random.choice(self.game.getActionSize(), p=vap)

    return a


class MonteCarlo():

  def __init__(self, game, nnet, args):
    """
    Args:
      game: OthelloGame instance
        Instance of the OthelloGame class above;
      nnet: OthelloNet instance
        Instance of the OthelloNNet class above;
      args: dictionary
        Instantiates number of iterations and episodes, controls temperature threshold, queue length,
        arena, checkpointing, and neural network parameters:
        learning-rate: 0.001, dropout: 0.3, epochs: 10, batch_size: 64,
        num_channels: 512
    """
    self.game = game
    self.nnet = nnet
    self.args = args

    self.Ps = {}  # Stores initial policy (returned by neural net)
    self.Es = {}  # Stores game.getGameEnded ended for board s

  # Call this rollout
  def simulate(self, canonicalBoard):
    """
    Simulate one Monte Carlo rollout

    Args:
      canonicalBoard: np.ndarray
        Canonical Board of size n x n [6x6 in this case]

    Returns:
      temp_v:
        Terminal State
    """
    s = self.game.stringRepresentation(canonicalBoard)
    init_start_state = s
    temp_v = 0
    isfirstAction = None
    current_player = -1  # opponent's turn (the agent has already taken an action before the simulation)
    self.Ps[s], _ = self.nnet.predict(canonicalBoard)

    for i in range(self.args.maxDepth):  # maxDepth

      if s not in self.Es:
        self.Es[s] = self.game.getGameEnded(canonicalBoard, 1)
      if self.Es[s] != 0:
        # Terminal state
        temp_v = self.Es[s] * current_player
        break

      self.Ps[s], v = self.nnet.predict(canonicalBoard)
      valids = self.game.getValidMoves(canonicalBoard, 1)
      self.Ps[s] = self.Ps[s] * valids  # Masking invalid moves
      sum_Ps_s = np.sum(self.Ps[s])

      if sum_Ps_s > 0:
        self.Ps[s] /= sum_Ps_s  # Renormalize
      else:
        # If all valid moves were masked make all valid moves equally probable
        # NB! All valid moves may be masked if either your NNet architecture is insufficient or you've get overfitting or something else.
        # If you have got dozens or hundreds of these messages you should pay attention to your NNet and/or training process.
        log.error("All valid moves were masked, doing a workaround.")
        self.Ps[s] = self.Ps[s] + valids
        self.Ps[s] /= np.sum(self.Ps[s])

      # Choose action according to the policy distribution
      a = np.random.choice(self.game.getActionSize(), p=self.Ps[s])
      # Find the next state and the next player
      next_s, next_player = self.game.getNextState(canonicalBoard, 1, a)
      canonicalBoard = self.game.getCanonicalForm(next_s, next_player)
      s = self.game.stringRepresentation(next_s)
      current_player *= -1
      # Initial policy
      self.Ps[s], v = self.nnet.predict(canonicalBoard)
      temp_v = v.item() * current_player

    return temp_v


class MonteCarloBasedPlayer():
  """
  Simulate Player based on Monte Carlo Algorithm
  """

  def __init__(self, game, nnet, args):
    """
    Args:
      game: OthelloGame instance
        Instance of the OthelloGame class above;
      nnet: OthelloNet instance
        Instance of the OthelloNNet class above;
      args: dictionary
        Instantiates number of iterations and episodes, controls temperature threshold, queue length,
        arena, checkpointing, and neural network parameters:
        learning-rate: 0.001, dropout: 0.3, epochs: 10, batch_size: 64,
        num_channels: 512
    """
    self.game = game
    self.nnet = nnet
    self.args = args
    self.mc = MonteCarlo(game, nnet, args)
    self.K = self.args.mc_topk

  def play(self, canonicalBoard):
    """
    Simulate Play on Canonical Board

    Args:
      canonicalBoard: np.ndarray
        Canonical Board of size n x n [6x6 in this case]

    Returns:
      best_action: tuple
        (avg_value, action) i.e., Average value associated with corresponding action
        i.e., Action with the highest topK probability
    """
    self.qsa = []
    s = self.game.stringRepresentation(canonicalBoard)
    Ps, v = self.nnet.predict(canonicalBoard)
    valids = self.game.getValidMoves(canonicalBoard, 1)
    Ps = Ps * valids  # Masking invalid moves
    sum_Ps_s = np.sum(Ps)

    if sum_Ps_s > 0:
      Ps /= sum_Ps_s  # Renormalize
    else:
      # If all valid moves were masked make all valid moves equally probable
      # NB! All valid moves may be masked if either your NNet architecture is insufficient or you've get overfitting or something else.
      # If you have got dozens or hundreds of these messages you should pay attention to your NNet and/or training process.
      log = logging.getLogger(__name__)
      log.error("All valid moves were masked, doing a workaround.")
      Ps = Ps + valids
      Ps /= np.sum(Ps)

    num_valid_actions = np.shape(np.nonzero(Ps))[1]

    if num_valid_actions < self.K:
      top_k_actions = np.argpartition(Ps,-num_valid_actions)[-num_valid_actions:]
    else:
      top_k_actions = np.argpartition(Ps,-self.K)[-self.K:]  # To get actions that belongs to top k prob

    for action in top_k_actions:
      next_s, next_player = self.game.getNextState(canonicalBoard, 1, action)
      next_s = self.game.getCanonicalForm(next_s, next_player)

      values = []

      # Do some rollouts
      for rollout in range(self.args.numMCsims):
        value = self.mc.simulate(next_s)
        values.append(value)

      # Average out values
      avg_value = np.mean(values)
      self.qsa.append((avg_value, action))

    self.qsa.sort(key=lambda a: a[0])
    self.qsa.reverse()
    best_action = self.qsa[0][1]
    return best_action

  def getActionProb(self, canonicalBoard, temp=1):
    """
    Get probabilities associated with each action

    Args:
      canonicalBoard: np.ndarray
        Canonical Board of size n x n [6x6 in this case]
      temp: Integer
        Signifies if game is in terminal state

    Returns:
      action_probs: List
        Probability associated with corresponding action
    """
    if self.game.getGameEnded(canonicalBoard, 1) != 0:
      return np.zeros((self.game.getActionSize()))

    else:
      action_probs = np.zeros((self.game.getActionSize()))
      best_action = self.play(canonicalBoard)
      action_probs[best_action] = 1

    return action_probs

The hyperparameters used throughout the notebook.

args = dotdict({
    'numIters': 1,  # In training, number of iterations = 1000 and num of episodes = 100
    'numEps': 1,  # Number of complete self-play games to simulate during a new iteration.
    'tempThreshold': 15,  # To control exploration and exploitation
    'updateThreshold': 0.6,  # During arena playoff, new neural net will be accepted if threshold or more of games are won.
    'maxlenOfQueue': 200,  # Number of game examples to train the neural networks.
    'numMCTSSims': 15,  # Number of games moves for MCTS to simulate.
    'arenaCompare': 10,  # Number of games to play during arena play to determine if new net will be accepted.
    'cpuct': 1,
    'maxDepth':5,  # Maximum number of rollouts
    'numMCsims': 5,  # Number of monte carlo simulations
    'mc_topk': 3,  # Top k actions for monte carlo rollout

    'checkpoint': './temp/',
    'load_model': False,
    'load_folder_file': ('/dev/models/8x100x50','best.pth.tar'),
    'numItersForTrainExamplesHistory': 20,

    # Define neural network arguments
    'lr': 0.001,  # learning rate
    'dropout': 0.3,
    'epochs': 10,
    'batch_size': 64,
    'device': DEVICE,
    'num_channels': 512,
})

Section 1: Plan using Monte Carlo Tree Search (MCTS)#

Time estimate: ~30 mins

Goal: Teach students to understand the core ideas behind Monte Carlo Tree Search (MCTS).

Video 1: Plan with MCTS#

Submit your feedback#

Coding Exercise 1: MCTS planner#

In building the MCTS planner, we will focus on the action selection part, particularly the objective function used. MCTS will use a combination of the current action-value function \(Q\) and the policy prior as follows:

(125)#\[\begin{equation} \underset{a}{\operatorname{argmax}} (Q(s_t, a)+u(s_t, a)) \end{equation}\]

with \(u(s_t, a)=c_{puct} \cdot P(s,a) \cdot \frac{\sqrt{\sum_b N(s,b)}}{1+N(s,a)}\). This effectively implements an Upper Confidence bound applied to Trees (UCT). UCT balances exploration and exploitation by taking the values stored from the MCTS into account. The trade-off is parametrized by \(c_{puct}\).

Note: Polynomial Upper Confidence Trees (PUCT) is the technical term for the alorithm below in which we sequentially run MCTS and store/use information from previous runs to explore and find optimal actions).

Exercise:

Finish the MCTS planner by using UCT to select actions to build the tree.
Deploy the MCTS planner to build a tree search for a given board position, producing value estimates and action counts for that position.

class MCTS():

  def __init__(self, game, nnet, args):
    """
    Args:
      game: OthelloGame instance
        Instance of the OthelloGame class above;
      nnet: OthelloNet instance
        Instance of the OthelloNNet class above;
      args: dictionary
        Instantiates number of iterations and episodes, controls temperature threshold, queue length,
        arena, checkpointing, and neural network parameters:
        learning-rate: 0.001, dropout: 0.3, epochs: 10, batch_size: 64,
        num_channels: 512
    """
    self.game = game
    self.nnet = nnet
    self.args = args
    self.Qsa = {}  # Stores Q values for s,a (as defined in the paper)
    self.Nsa = {}  # Stores #times edge s,a was visited
    self.Ns = {}  # Stores #times board s was visited
    self.Ps = {}  # Stores initial policy (returned by neural net)
    self.Es = {}  # Stores game.getGameEnded ended for board s
    self.Vs = {}  # Stores game.getValidMoves for board s

  def search(self, canonicalBoard):
    """
    Perform one iteration of MCTS.

    It is recursively called till a leaf node is found. The action chosen at
    each node is one that has the maximum upper confidence bound.
    Once a leaf node is found, the neural network is called to return an
    initial policy P and a value v for the state. This value is propagated
    up the search path. In case the leaf node is a terminal state, the
    outcome is propagated up the search path. The values of Ns, Nsa, Qsa are
    updated.
    NOTE: the return values are the negative of the value of the current
    state. This is done since v is in [-1,1] and if v is the value of a
    state for the current player, then its value is -v for the other player.

    Args:
      canonicalBoard: np.ndarray
        Canonical Board of size n x n [6x6 in this case]

    Returns:
        v: Float
          The negative of the value of the current canonicalBoard
    """
    s = self.game.stringRepresentation(canonicalBoard)

    if s not in self.Es:
      self.Es[s] = self.game.getGameEnded(canonicalBoard, 1)
    if self.Es[s] != 0:
      # Terminal node
      return -self.Es[s]

    if s not in self.Ps:
      # Leaf node
      self.Ps[s], v = self.nnet.predict(canonicalBoard)
      valids = self.game.getValidMoves(canonicalBoard, 1)
      self.Ps[s] = self.Ps[s] * valids  # Masking invalid moves
      sum_Ps_s = np.sum(self.Ps[s])
      if sum_Ps_s > 0:
        self.Ps[s] /= sum_Ps_s  # Renormalize
      else:
        # If all valid moves were masked make all valid moves equally probable
        # NB! All valid moves may be masked if either your NNet architecture is
        # insufficient or you've get overfitting or something else.
        # If you have got dozens or hundreds of these messages you should
        # pay attention to your NNet and/or training process.
        log = logging.getLogger(__name__)
        log.error("All valid moves were masked, doing a workaround.")
        self.Ps[s] = self.Ps[s] + valids
        self.Ps[s] /= np.sum(self.Ps[s])

      self.Vs[s] = valids
      self.Ns[s] = 0

      return -v

    valids = self.Vs[s]
    cur_best = -float('inf')
    best_act = -1

    ############################################################################
    ## TODO for students:
    #  Implement the highest upper confidence bound depending whether we observed
    #  the state-action pair which is stored in self.Qsa[(s, a)].
    #  You can find the formula in the slide 52 in video 8 above.
    #  Fill out function and remove
    raise NotImplementedError("Complete the for loop")
    ############################################################################
    # Pick the action with the highest upper confidence bound
    for a in range(self.game.getActionSize()):
      if valids[a]:
        if (s, a) in self.Qsa:
          u = ... + ... * ... * math.sqrt(...) / (1 + ...)
        else:
          u = ... * ... * math.sqrt(... + 1e-8)

        if u > cur_best:
          cur_best = u
          best_act = a

    a = best_act
    next_s, next_player = self.game.getNextState(canonicalBoard, 1, a)
    next_s = self.game.getCanonicalForm(next_s, next_player)

    v = self.search(next_s)

    if (s, a) in self.Qsa:
      self.Qsa[(s, a)] = (self.Nsa[(s, a)] * self.Qsa[(s, a)] + v) / (self.Nsa[(s, a)] + 1)
      self.Nsa[(s, a)] += 1

    else:
      self.Qsa[(s, a)] = v
      self.Nsa[(s, a)] = 1

    self.Ns[s] += 1
    return -v

  def getNsa(self):
    return self.Nsa

Click for solution

Submit your feedback#

Section 2: Use MCTS to play games#

Time estimate: ~10 mins

Goal: Learn how to use the results of MCTS to play games.

Exercise:

Plug the MCTS planner into an agent.
Play games against other agents.
Explore the contributions of prior network, value function, number of simulations/time to play and explore/exploit parameters.

Video 2: Play with MCTS#

Submit your feedback#

Coding Exercise 2: Agent that uses an MCTS planner#

Now we can use the MCTS planner and play the game! We will again let the MCTS planner play against players with other policies.

# Load MCTS model from the repository
mcts_model_save_name = 'MCTS.pth.tar'
path = "nma_rl_games/alpha-zero/pretrained_models/models/"

class MonteCarloTreeSearchBasedPlayer():

  def __init__(self, game, nnet, args):
    """
    Args:
      game: OthelloGame instance
        Instance of the OthelloGame class above;
      nnet: OthelloNet instance
        Instance of the OthelloNNet class above;
      args: dictionary
        Instantiates number of iterations and episodes, controls temperature threshold, queue length,
        arena, checkpointing, and neural network parameters:
        learning-rate: 0.001, dropout: 0.3, epochs: 10, batch_size: 64,
        num_channels: 512
    """
    self.game = game
    self.nnet = nnet
    self.args = args
    self.mcts = MCTS(game, nnet, args)

  def play(self, canonicalBoard, temp=1):
    """
    Args:
      canonicalBoard: np.ndarray
        Canonical Board of size n x n [6x6 in this case]
      temp: Integer
        Signifies if game is in terminal state

    Returns:
      List of probabilities for all actions if temp is 0
      Best action based on max probability otherwise
    """
    for i in range(self.args.numMCTSSims):

      ##########################################################################
      ## TODO for students:
      #  Run MCTS search function.
      #  Fill out function and remove
      raise NotImplementedError("Plug the planner")
      ##########################################################################
      ...

    s = self.game.stringRepresentation(canonicalBoard)
    ############################################################################
    ## TODO for students:
    #  Call the Nsa function from MCTS class and store it in the self.Nsa
    #  Fill out function and remove
    raise NotImplementedError("Compute Nsa (number of times edge s,a was visited)")
    ############################################################################
    self.Nsa = ...
    self.counts = [self.Nsa[(s, a)] if (s, a) in self.Nsa else 0 for a in range(self.game.getActionSize())]

    if temp == 0:
      bestAs = np.array(np.argwhere(self.counts == np.max(self.counts))).flatten()
      bestA = np.random.choice(bestAs)
      probs = [0] * len(self.counts)
      probs[bestA] = 1
      return probs

    self.counts = [x ** (1. / temp) for x in self.counts]
    self.counts_sum = float(sum(self.counts))
    probs = [x / self.counts_sum for x in self.counts]
    return np.argmax(probs)

  def getActionProb(self, canonicalBoard, temp=1):
    """
    Args:
      canonicalBoard: np.ndarray
        Canonical Board of size n x n [6x6 in this case]
      temp: Integer
        Signifies if game is in terminal state

    Returns:
      action_probs: List
        Probability associated with corresponding action
    """
    action_probs = np.zeros((self.game.getActionSize()))
    best_action = self.play(canonicalBoard)
    action_probs[best_action] = 1

    return action_probs


set_seed(seed=SEED)
game = OthelloGame(6)
rp = RandomPlayer(game).play  # All players
num_games = 20  # Games
n1 = NNet(game)  # nnet players
n1.load_checkpoint(folder=path, filename=mcts_model_save_name)
args1 = dotdict({'numMCTSSims': 50, 'cpuct':1.0})

## Uncomment below to check your agent!
# print('\n******MCTS player versus random player******')
# mcts1 = MonteCarloTreeSearchBasedPlayer(game, n1, args1)
# n1p = lambda x: np.argmax(mcts1.getActionProb(x, temp=0))
# arena = Arena.Arena(n1p, rp, game, display=OthelloGame.display)
# MCTS_result = arena.playGames(num_games, verbose=False)
# print(f"\nNumber of games won by player1 = {MCTS_result[0]}, "
#       f"number of games won by player2 = {MCTS_result[1]}, out of {num_games} games")
# win_rate_player1 = MCTS_result[0]/num_games
# print(f"\nWin rate for player1 over {num_games} games: {round(win_rate_player1*100, 1)}%")

Number of games won by player1 = 19, num of games won by player2 = 1, out of 20 games

Win rate for player1 over 20 games: 95.0%

Click for solution

Load in trained value and policy networks#

Random seed 2023 has been set.
Random seed 2023 has been set.

MCTS player against Value-based player#

print('\n******MCTS player versus value-based player******')
set_seed(seed=SEED)
vp = ValueBasedPlayer(game, vnet).play  # Value-based player
arena = Arena.Arena(n1p, vp, game, display=OthelloGame.display)
MC_result = arena.playGames(num_games, verbose=False)

print(f"\nNumber of games won by player1 = {MC_result[0]}, "
      f"number of games won by player2 = {MC_result[1]}, out of {num_games} games")
win_rate_player1 = MC_result[0]/num_games
print(f"\nWin rate for player1 over {num_games} games: {round(win_rate_player1*100, 1)}%")

Number of games won by player1 = 17, number of games won by player2 = 3, out of 20 games

Win rate for player1 over 20 games: 85.0%

MCTS player against Policy-based player#

print('\n******MCTS player versus policy-based player******')
set_seed(seed=SEED)
pp = PolicyBasedPlayer(game, pnet).play  # Policy-based player
arena = Arena.Arena(n1p, pp, game, display=OthelloGame.display)
MC_result = arena.playGames(num_games, verbose=False)

print(f"\nNumber of games won by player1 = {MC_result[0]}, "
      f"number of games won by player2 = {MC_result[1]}, out of {num_games} games")
win_rate_player1 = MC_result[0]/num_games
print(f"\nWin rate for player1 over {num_games} games: {round(win_rate_player1*100, 1)}%")

Number of games won by player1 = 20, number of games won by player2 = 0, out of 20 games

Win rate for player1 over 20 games: 100.0%

MCTS player against Monte-Carlo player#

mc_model_save_name = 'MC.pth.tar'
path = "nma_rl_games/alpha-zero/pretrained_models/models/"

n2 = NNet(game)  # nNet players
n2.load_checkpoint(folder=path, filename=mc_model_save_name)
args2 = dotdict({'numMCsims': 10, 'maxRollouts':5, 'maxDepth':5, 'mc_topk': 3})

print('\n******MCTS player versus MC player******')
set_seed(seed=SEED)
mc = MonteCarloBasedPlayer(game, n2, args2)
n2p = lambda x: np.argmax(mc.getActionProb(x))
arena = Arena.Arena(n1p, n2p, game, display=OthelloGame.display)
MC_result = arena.playGames(num_games, verbose=False)

print(f"\nNumber of games won by player1 = {MC_result[0]}, "
      f"number of games won by player2 = {MC_result[1]}, out of {num_games} games")
win_rate_player1 = MC_result[0]/num_games
print(f"\nWin rate for player1 over {num_games} games: {round(win_rate_player1*100, 1)}%")

Number of games won by player1 = 16, number of games won by player2 = 4, out of 20 games

Win rate for player1 over 20 games: 80.0%

Submit your feedback#

Summary#

In this tutorial, you have learned about players with Monte Carlo Tree Search planner and compared them to random, value-based, policy-based, and Monte-Carlo players.

Bonus Tutorial: Planning with Monte Carlo Tree Search

Contents

Bonus Tutorial: Planning with Monte Carlo Tree Search#

Tutorial Objectives#

Setup#

Install dependencies#

Install and import feedback gadget#

Set random seed#

Set device (GPU or CPU). Execute set_device()#

Download the modules#

Helper functions from previous tutorials#

Section 1: Plan using Monte Carlo Tree Search (MCTS)#

Video 1: Plan with MCTS#

Submit your feedback#

Coding Exercise 1: MCTS planner#

Submit your feedback#

Section 2: Use MCTS to play games#

Video 2: Play with MCTS#

Submit your feedback#

Coding Exercise 2: Agent that uses an MCTS planner#

Load in trained value and policy networks#

MCTS player against Value-based player#

MCTS player against Policy-based player#

MCTS player against Monte-Carlo player#

Submit your feedback#

Summary#

Set device (GPU or CPU). Execute `set_device()`#