def test_insert_some_coins():
b = Board()
assert b.turn() == 'O'
b = b.insert(3)
assert b.turn() == 'X'
assert b == Board([0, 0, 0, 0b01, 0, 0, 0])
assert b.valid_actions() == (0, 1, 2, 3, 4, 5, 6)
b = b.insert(2)
assert b.turn() == 'O'
assert b == Board([0, 0, 0b10, 0b01, 0, 0, 0])
assert b.valid_actions() == (0, 1, 2, 3, 4, 5, 6)
b = b.insert(2)
assert b.turn() == 'X'
assert b == Board([0, 0, 0b0110, 0b01, 0, 0, 0])
assert b.valid_actions() == (0, 1, 2, 3, 4, 5, 6)
b = b.insert(2)
assert b.turn() == 'O'
assert b == Board([0, 0, 0b100110, 0b01, 0, 0, 0])
assert b.valid_actions() == (0, 1, 2, 3, 4, 5, 6)
b = b.insert(2)
assert b.turn() == 'X'
assert b == Board([0, 0, 0b01100110, 0b01, 0, 0, 0])
assert b.valid_actions() == (0, 1, 2, 3, 4, 5, 6)
b = b.insert(2)
assert b.turn() == 'O'
assert b == Board([0, 0, 0b1001100110, 0b01, 0, 0, 0])
assert b.valid_actions() == (0, 1, 2, 3, 4, 5, 6)
b = b.insert(2)
assert b.turn() == 'X'
assert b == Board([0, 0, 0b011001100110, 0b01, 0, 0, 0])
assert b.valid_actions() == (0, 1, 3, 4, 5, 6)
def select_action(policy, board: Board, cuda=False, noise=0):
# Get probabilities from neural network
state = torch.from_numpy(board.matrix().reshape(BOARD_ROWS * BOARD_COLS)).float().unsqueeze(0)
if cuda:
state = state.cuda()
probs = policy(Variable(state))
# Exclude any results that are not allowed
mult_np = np.zeros(len(POSSIBLE_ACTIONS), dtype=np.float32)
allowed_actions = board.valid_actions()
for i in POSSIBLE_ACTIONS:
if i in allowed_actions:
mult_np[i] = 1
# Always choose winning move
for a in allowed_actions:
hypothetical_board = board.insert(a)
if hypothetical_board.winner() == board.turn():
mult_np = np.zeros(len(POSSIBLE_ACTIONS), dtype=np.float32)
mult_np[a] = 1
mult = Variable(torch.from_numpy(mult_np))
noise = Variable(torch.from_numpy(mult_np * noise))
if cuda:
mult = mult.cuda()
noise = noise.cuda()
probs = probs * mult + noise
if torch.sum(probs * mult).data[0] < 1e-40:
# Neural network only offered things that are not allowed, so we go for random
probs = probs + mult
return probs.multinomial()