def train(qTables, numGames, alpha, tryHard):
tryHardGrowth = (1 - tryHard) / numGames
for i in tqdm(range(numGames)):
board = tt.genBoard()
movesLeft = True
winner = False
player = 2
keysSoFar = []
movesSoFar = []
computersPlayer = random.randint(1, 2)
while (movesLeft and not winner):
if player == computersPlayer:
bestMove = pickBestNextMove(qTables, keysSoFar, board, player,
computersPlayer, tryHard)
movesSoFar.append(bestMove)
tt.applyMove(player, bestMove, board)
else:
moves = tt.listEmpties(board)
randomMove = random.choice(moves)
tt.applyMove(player, randomMove, board)
player = tt.togglePlayer(player)
winner = tt.getWinner(board)
movesLeft = not tt.noMoreMoves(board)
score = scoreEndBoard(board, winner, computersPlayer)
updateQTable(score, qTables, keysSoFar, movesSoFar, alpha)
tryHard = tryHard + tryHardGrowth
def test(net, epochs):
net.eval()
numWins = 0
numLosses = 0
numTies = 0
for i in tqdm(range(epochs)):
player = 2
computersPlayer = random.randint(1,2)
board = np.zeros(shape = (3, 3))
# board = np.random.randint(low = 0, high = 3, size = (3, 3))
movesLeft = np.any(np.where(board == 0, 1, 0))
winner = tt.getWinner(board)
while(not winner and movesLeft):
if player == computersPlayer:
# generate a move
oneHot = oneHotTicTacToe(board, computersPlayer).view(1, 1, 18)
output = net(oneHot)
# mask out invalid moves
invalidMoves = np.where( board.flatten() > 0, True, False)
maskedOutput = output.clone().view(9)
maskedOutput[invalidMoves] = -10
values, index = maskedOutput.max(0)
# apply the move
move = index
board = board.flatten()
board[move] = computersPlayer
board = board.reshape(3, 3)
else: # opponents turn
empties = tt.listEmpties(board)
randomMove = random.choice(empties)
tt.applyMove(player, randomMove, board)
player = tt.togglePlayer(player)
movesLeft = np.any(np.where(board == 0, 1, 0))
winner = tt.getWinner(board)
if winner == computersPlayer:
numWins += 1
elif winner == tt.togglePlayer(computersPlayer):
numLosses += 1
else: # winner == False
numTies += 1
return numWins, numLosses, numTies
def test(mct, numGames, numSims):
numWins = 0
numTies = 0
numLosses = 0
for i in tqdm(range(numGames)):
board = tt.genBoard()
movesLeft = True
winner = False
player = 2
computersPlayer = random.randint(1, 2)
while (movesLeft and not winner):
if player == computersPlayer:
simulateChildren(mct, board, player, computersPlayer, numSims)
bestBoard = pickBestNextMove(mct, board, player)
# print("################")
# tt.printBoard(board)
# tt.printBoard(bestBoard)
# print("BESTMOVE")
board = bestBoard
else:
moves = tt.listEmpties(board)
randomMove = random.choice(moves)
tt.applyMove(player, randomMove, board)
player = tt.togglePlayer(player)
winner = tt.getWinner(board)
movesLeft = not tt.noMoreMoves(board)
if winner == computersPlayer:
numWins += 1
elif winner == tt.togglePlayer(computersPlayer):
numLosses += 1
else: # tie
numTies += 1
return numWins, numLosses, numTies
def simulate(numSimulations, board, player, myPlayer):
originBoard = copy.deepcopy(board)
originPlayer = player
totalScore = 0
for i in range(numSimulations):
simBoard = copy.deepcopy(originBoard)
simPlayer = originPlayer
winner = tt.getWinner(simBoard)
movesLeft = not tt.noMoreMoves(simBoard)
while (movesLeft and not winner):
moves = tt.listEmpties(simBoard)
randomMove = random.choice(moves)
tt.applyMove(simPlayer, randomMove, simBoard)
simPlayer = tt.togglePlayer(simPlayer)
winner = tt.getWinner(simBoard)
movesLeft = not tt.noMoreMoves(simBoard)
score = scoreEndBoard(simBoard, winner, myPlayer)
totalScore += score
return totalScore
def test(qTables, numGames, tryHard=1.0):
numWins = 0
numTies = 0
numLosses = 0
for i in tqdm(range(numGames)):
board = tt.genBoard()
movesLeft = True
winner = False
player = 2
keysSoFar = []
movesSoFar = []
computersPlayer = random.randint(1, 2)
while (movesLeft and not winner):
if player == computersPlayer:
bestMove = pickBestNextMove(qTables, keysSoFar, board, player,
computersPlayer, tryHard)
movesSoFar.append(bestMove)
tt.applyMove(player, bestMove, board)
else:
moves = tt.listEmpties(board)
randomMove = random.choice(moves)
tt.applyMove(player, randomMove, board)
player = tt.togglePlayer(player)
winner = tt.getWinner(board)
movesLeft = not tt.noMoreMoves(board)
if winner == computersPlayer:
numWins += 1
elif winner == tt.togglePlayer(computersPlayer):
numLosses += 1
else: # tie
numTies += 1
return numWins, numLosses, numTies
def playGame():
saveQTables = False
fileName = 'qTables.pickle'
qTables = {}
keysSoFar = []
movesSoFar = []
tryHard = 0
alpha = 0.9
numTrials = 1000000
if saveQTables:
train(qTables, numTrials, alpha, tryHard)
f = open(fileName, 'wb')
pickle.dump(qTables, f, pickle.HIGHEST_PROTOCOL)
f.close()
else:
f = open(fileName, 'rb')
qTables = pickle.load(f)
f.close()
numWins, numLosses, numTies = test(qTables, 1000)
print("VS RANDOM OPPONENT...")
print("numWins:" + str(numWins))
print("numLosses:" + str(numLosses))
print("numTies:" + str(numTies))
quit()
board = tt.genBoard()
movesLeft = True
winner = False
player = 2
computersPlayer = random.randint(1, 2)
print("NEW GAME")
if computersPlayer == 2:
print("COMPUTER GOES FIRST...")
while (movesLeft and not winner):
if player == 2:
print("X's Turn")
else: # player == 1
print("O's Turn")
tt.printBoard(board)
if player == computersPlayer:
bestMove = pickBestNextMove(qTables,
keysSoFar,
board,
player,
computersPlayer,
tryHard=1.0,
verbose=True)
movesSoFar.append(bestMove)
tt.applyMove(player, bestMove, board)
player = tt.togglePlayer(player)
elif player == tt.togglePlayer(computersPlayer):
validMove = False
while validMove == False:
move = input("input move of form 'y x' ")
y = int(move[0])
x = int(move[2])
# validate move
if board[y][x] is not 0:
print("!!!INVALID MOVE!!!")
continue
else:
validMove = True
board[y][x] = tt.togglePlayer(computersPlayer)
player = tt.togglePlayer(player)
winner = tt.getWinner(board)
movesLeft = not tt.noMoreMoves(board)
tt.printBoard(board)
score = scoreEndBoard(board, winner, computersPlayer)
updateQTable(score, qTables, keysSoFar, movesSoFar, alpha)
for key in keysSoFar:
pprint(key)
pprint(qTables[key])
if winner:
if winner == 2:
print("WINNER: X")
else: # winner == 1
print("WINNER: O")
else:
print("TIE")
def train(net, criterion, optimizer, epochs):
net.train()
for i in tqdm(range(epochs)):
player = 2
computersPlayer = random.randint(1,2)
optimizer.zero_grad()
board = np.zeros(shape = (3, 3))
movesLeft = np.any(np.where(board == 0, 1, 0))
winner = tt.getWinner(board)
gameDuration = 0
moves = []
outputs = []
while(not winner and movesLeft):
if player == computersPlayer:
# generate a move
oneHot = oneHotTicTacToe(board, computersPlayer).view(1, 1, 18)
output = net(oneHot)
# mask out invalid moves
invalidMoves = np.where( board.flatten() > 0, True, False)
maskedOutput = output.clone().view(9)
maskedOutput[invalidMoves] = -10
values, index = maskedOutput.max(0)
# apply the move
move = index
board = board.flatten()
board[move] = computersPlayer
board = board.reshape(3, 3)
# store for later
moves.append(move)
outputs.append(output)
else: # opponents turn
empties = tt.listEmpties(board)
randomMove = random.choice(empties)
tt.applyMove(player, randomMove, board)
player = tt.togglePlayer(player)
gameDuration += 1
movesLeft = np.any(np.where(board == 0, 1, 0))
winner = tt.getWinner(board)
# get end score of game
score = scoreEndBoard(board, winner, computersPlayer)
# gameDurationMultiplier = 1.0 - gameDuration / 10
# gameDurationMultiplier = gameDurationMultiplier * 0.9
dilutionFactor = 0.9
totalDilutant = 1.0
for i, move in reversed(list(enumerate(moves))):
totalDilutant *= dilutionFactor
output = outputs[i]
target = output.clone().view(9)
target[move] = score * totalDilutant
target = target.view(1, 1, 9)
optimizer.zero_grad()
loss = criterion(output, target)
loss.backward()
optimizer.step()
def test(net, criterion, optimizer, epochs):
numInvalidMoves = 0
numWins = 0
numLosses = 0
numTies = 0
optimizer.zero_grad()
for i in tqdm(range(epochs)):
player = 2
computersPlayer = random.randint(1,2)
board = np.zeros(shape = (3, 3))
# board = np.random.randint(low = 0, high = 3, size = (3, 3))
movesLeft = np.any(np.where(board == 0, 1, 0))
winner = tt.getWinner(board)
while(not winner and movesLeft):
if player == computersPlayer:
move = None
moveValid = False
while not moveValid:
# generate a move
oneHot = oneHotTicTacToe(board, computersPlayer).view(1, 1, 18)
output = net(oneHot)
values, index = output.view(9).max(0)
if board.flatten()[index] == 0: # if move is valid
moveValid = True
# apply the move
move = index
board = board.flatten()
board[move] = computersPlayer
board = board.reshape(3, 3)
else: # invalid move, prime the whip
# print("invalid move")
numInvalidMoves += 1
optimizer.zero_grad()
validMoves = np.where(board == 0, 1, 0)
target = torch.tensor(validMoves, dtype=torch.float).view(1, 1, 9)
loss = criterion(output, target)
loss.backward()
optimizer.step()
else: # opponents turn
empties = tt.listEmpties(board)
randomMove = random.choice(empties)
tt.applyMove(player, randomMove, board)
player = tt.togglePlayer(player)
movesLeft = np.any(np.where(board == 0, 1, 0))
winner = tt.getWinner(board)
if winner == computersPlayer:
numWins += 1
elif winner == tt.togglePlayer(computersPlayer):
numLosses += 1
else: # winner == False
numTies += 1
return numWins, numLosses, numTies
def train(net, criterion, optimizer, epochs):
numInvalidMoves = 0
for i in tqdm(range(epochs)):
player = 2
computersPlayer = random.randint(1,2)
optimizer.zero_grad()
board = np.zeros(shape = (3, 3))
# board = np.random.randint(low = 0, high = 3, size = (3, 3))
movesLeft = np.any(np.where(board == 0, 1, 0))
winner = tt.getWinner(board)
moves = []
outputs = []
while(not winner and movesLeft):
if player == computersPlayer:
move = None
moveValid = False
while not moveValid:
# generate a move
oneHot = oneHotTicTacToe(board, computersPlayer).view(1, 1, 18)
output = net(oneHot)
values, index = output.view(9).max(0)
if board.flatten()[index] == 0: # if move is valid
moveValid = True
# apply the move
move = index
board = board.flatten()
board[move] = computersPlayer
board = board.reshape(3, 3)
# store for later
moves.append(move)
outputs.append(output)
else: # invalid move, prime the whip
# print("invalid move")
numInvalidMoves += 1
optimizer.zero_grad()
validMoves = np.where(board == 0, 1, 0)
target = torch.tensor(validMoves, dtype=torch.float).view(1, 1, 9)
loss = criterion(output, target)
loss.backward()
optimizer.step()
else: # opponents turn
empties = tt.listEmpties(board)
randomMove = random.choice(empties)
tt.applyMove(player, randomMove, board)
player = tt.togglePlayer(player)
movesLeft = np.any(np.where(board == 0, 1, 0))
winner = tt.getWinner(board)
# get end score of game
score = scoreEndBoard(board, winner, computersPlayer)
for i, move in enumerate(moves):
output = outputs[i]
target = output.clone().view(9)
target[move] = score
target = target.view(1, 1, 9)
optimizer.zero_grad()
loss = criterion(output, target)
loss.backward()
optimizer.step()