def run_and_get_new_weights(init_weights, win0, win1):
state = GameState(size=19)
policy = CNNPolicy.load_model(
os.path.join('test_data', 'minimodel.json'))
policy.model.set_weights(init_weights)
optimizer = BatchedReinforcementLearningSGD(lr=0.01, ng=2)
policy.model.compile(loss=log_loss, optimizer=optimizer)
# Make moves on the state and get trainable (state, action) pairs from them.
moves = [(2, 2), (16, 16), (3, 17), (16, 2), (4, 10), (10, 3)]
state_tensors = []
action_tensors = []
for m in moves:
(st_tensor,
mv_tensor) = _make_training_pair(state, m,
policy.preprocessor)
state_tensors.append(st_tensor)
action_tensors.append(mv_tensor)
state.do_move(m)
for i, (s, a) in enumerate(zip(state_tensors, action_tensors)):
# Put even state/action pairs in game 0, odd ones in game 1.
game_idx = i % 2
optimizer.set_current_game(game_idx)
is_last_move = i + 2 >= len(moves)
if is_last_move:
if game_idx == 0:
optimizer.set_result(game_idx, win0)
else:
optimizer.set_result(game_idx, win1)
# train_on_batch accumulates gradients, and should only cause a change to parameters
# on the first call after the final set_result() call
policy.model.train_on_batch(s, a)
return policy.model.get_weights()
def test_positional_superko(self):
move_list = [(0, 3), (0, 4), (1, 3), (1, 4), (2, 3), (2, 4), (2, 2),
(3, 4), (2, 1), (3, 3), (3, 1), (3, 2), (3, 0), (4, 2),
(1, 1), (4, 1), (8, 0), (4, 0), (8, 1), (0, 2), (8, 2),
(0, 1), (8, 3), (1, 0), (8, 4), (2, 0), (0, 0)]
# 0 1 2 3 4 5 6 7 8 9
# 0 . W W B W . . . . .
# 1 W B . B W . . . . .
# 2 W B B B W . . . . .
# 3 B B W W W . . . . .
# 4 W W W . . . . . . .
# 5 . . . . . . . . . .
# 6 . . . . . . . . . .
# 7 . . . . . . . . . .
# 8 B B B B B . . . . .
# 9 . . . . . . . . . .
gs = GameState(size=9)
for move in move_list:
gs.do_move(move)
self.assertTrue(gs.is_legal((1, 0)))
gs = GameState(size=9, enforce_superko=True)
for move in move_list:
gs.do_move(move)
self.assertFalse(gs.is_legal((1, 0)))
def test_output_size(self):
policy19 = CNNPolicy(["board", "liberties", "sensibleness", "capture_size"], board=19)
output = policy19.forward(policy19.preprocessor.state_to_tensor(GameState(19)))
self.assertEqual(output.shape, (1, 19 * 19))
policy13 = CNNPolicy(["board", "liberties", "sensibleness", "capture_size"], board=13)
output = policy13.forward(policy13.preprocessor.state_to_tensor(GameState(13)))
self.assertEqual(output.shape, (1, 13 * 13))
def test_probabilistic_player(self):
gs = GameState()
policy = CNNPolicy(["board", "ones", "turns_since"])
player = ProbabilisticPolicyPlayer(policy)
for i in range(20):
move = player.get_move(gs)
self.assertIsNotNone(move)
gs.do_move(move)
def testApplyAndResetOnGamesFinished(self):
policy = CNNPolicy.load_model(
os.path.join('test_data', 'minimodel.json'))
state = GameState(size=19)
optimizer = BatchedReinforcementLearningSGD(lr=0.01, ng=2)
policy.model.compile(loss=log_loss, optimizer=optimizer)
# Helper to check initial conditions of the optimizer.
def assertOptimizerInitialConditions():
for v in optimizer.gradient_sign:
self.assertEqual(K.eval(v), 0)
self.assertEqual(K.eval(optimizer.running_games), 2)
initial_parameters = policy.model.get_weights()
def assertModelEffect(changed):
any_change = False
for cur, init in zip(policy.model.get_weights(),
initial_parameters):
if not np.allclose(init, cur):
any_change = True
break
self.assertEqual(any_change, changed)
assertOptimizerInitialConditions()
# Make moves on the state and get trainable (state, action) pairs from them.
state_tensors = []
action_tensors = []
moves = [(2, 2), (16, 16), (3, 17), (16, 2), (4, 10), (10, 3)]
for m in moves:
(st_tensor,
mv_tensor) = _make_training_pair(state, m, policy.preprocessor)
state_tensors.append(st_tensor)
action_tensors.append(mv_tensor)
state.do_move(m)
for i, (s, a) in enumerate(zip(state_tensors, action_tensors)):
# Even moves in game 0, odd moves in game 1
game_idx = i % 2
optimizer.set_current_game(game_idx)
is_last_move = i + 2 >= len(moves)
if is_last_move:
# Mark game 0 as a win and game 1 as a loss.
optimizer.set_result(game_idx, game_idx == 0)
else:
# Games not finished yet; assert no change to optimizer state.
assertOptimizerInitialConditions()
# train_on_batch accumulates gradients, and should only cause a change to parameters
# on the first call after the final set_result() call
policy.model.train_on_batch(s, a)
if i + 1 < len(moves):
assertModelEffect(changed=False)
else:
assertModelEffect(changed=True)
# Once both games finished, the last call to train_on_batch() should have triggered a reset
# to the optimizer parameters back to initial conditions.
assertOptimizerInitialConditions()
def test_eye_recursion(self):
# a checkerboard pattern of black is 'technically' all true eyes
# mutually supporting each other
gs = GameState(7)
for x in range(gs.size):
for y in range(gs.size):
if (x + y) % 2 == 1:
gs.do_move((x, y), go.BLACK)
self.assertTrue(gs.is_eye((0, 0), go.BLACK))
def test_snapback_is_not_ko(self):
gs = GameState(size=5)
# B X W B .
# W W B . .
# . . . . .
# . . . . .
# . . . . .
# imagine black plays at 'X' capturing the white stone at (2, 0).
# White may play again at (2, 0) to capture the black stones
# at (0, 0), (1, 0). this is a 'snapback' not 'ko'
# since it doesn't return the game to a previous position
B = [(0, 0), (2, 1), (3, 0)]
W = [(0, 1), (1, 1), (2, 0)]
for (b, w) in zip(B, W):
gs.do_move(b)
gs.do_move(w)
# do the capture of the single white stone
gs.do_move((1, 0))
# there should be no ko
self.assertIsNone(gs.ko)
self.assertTrue(gs.is_legal((2, 0)))
# now play the snapback
gs.do_move((2, 0))
# check that the numbers worked out
self.assertEqual(gs.num_black_prisoners, 2)
self.assertEqual(gs.num_white_prisoners, 1)
def test_sensible_greedy(self):
gs = GameState()
policy = CNNPolicy(["board", "ones", "turns_since"])
player = GreedyPolicyPlayer(policy)
empty = (10, 10)
for x in range(19):
for y in range(19):
if (x, y) != empty:
gs.do_move((x, y), go.BLACK)
gs.current_player = go.BLACK
self.assertIsNone(player.get_move(gs))
def test_copy_maintains_shared_sets(self):
gs = GameState(7)
gs.do_move((4, 4), go.BLACK)
gs.do_move((4, 5), go.BLACK)
# assert that gs has *the same object* referenced by group/liberty sets
self.assertTrue(gs.group_sets[4][5] is gs.group_sets[4][4])
self.assertTrue(gs.liberty_sets[4][5] is gs.liberty_sets[4][4])
gs_copy = gs.copy()
self.assertTrue(gs_copy.group_sets[4][5] is gs_copy.group_sets[4][4])
self.assertTrue(
gs_copy.liberty_sets[4][5] is gs_copy.liberty_sets[4][4])
def parse(boardstr):
'''Parses a board into a gamestate, and returns the location of any moves
marked with anything other than 'B', 'X', '#', 'W', 'O', or '.'
Rows are separated by '|', spaces are ignored.
'''
boardstr = boardstr.replace(' ', '')
board_size = max(boardstr.index('|'), boardstr.count('|'))
st = GameState(size=board_size)
moves = {}
for row, rowstr in enumerate(boardstr.split('|')):
for col, c in enumerate(rowstr):
if c == '.':
continue # ignore empty spaces
elif c in 'BX#':
st.do_move((row, col), color=BLACK)
elif c in 'WO':
st.do_move((row, col), color=WHITE)
else:
# move reference
assert c not in moves, "{} already used as a move marker".format(
c)
moves[c] = (row, col)
return st, moves
def setUp(self):
# 0 1 2 3 4 5 6 7 8 9 A B
# 0 . . . . . . . . . . . .
# 1 . . . . . . . . . . . .
# 2 . . . . . . . . . . . .
# 3 . . . . . . . . . . . .
# 4 . . . . . B B . . . . .
# 5 . . . . . W B . . . . .
# 6 . . . . . . B . . . . .
# 7 . . . . . . . . . . . .
# 8 . . . . . . . . . . . .
# 9 . . . . . . . . . . W .
# A . . . . . . . . . . W W
# B . . . . . . . . . . . .
self.s = GameState()
self.s.do_move((4, 5))
self.s.do_move((5, 5))
self.s.do_move((5, 6))
self.s.do_move((10, 10))
self.s.do_move((4, 6))
self.s.do_move((10, 11))
self.s.do_move((6, 6))
self.s.do_move((9, 10))
def test_liberties_after_capture(self):
# creates 3x3 black group in the middle, that is then all captured
# ...then an assertion is made that the resulting liberties after
# capture are the same as if the group had never been there
gs_capture = GameState(7)
gs_reference = GameState(7)
# add in 3x3 black stones
for x in range(2, 5):
for y in range(2, 5):
gs_capture.do_move((x, y), go.BLACK)
# surround the black group with white stones
# and set the same white stones in gs_reference
for x in range(2, 5):
gs_capture.do_move((x, 1), go.WHITE)
gs_capture.do_move((x, 5), go.WHITE)
gs_reference.do_move((x, 1), go.WHITE)
gs_reference.do_move((x, 5), go.WHITE)
gs_capture.do_move((1, 1), go.WHITE)
gs_reference.do_move((1, 1), go.WHITE)
for y in range(2, 5):
gs_capture.do_move((1, y), go.WHITE)
gs_capture.do_move((5, y), go.WHITE)
gs_reference.do_move((1, y), go.WHITE)
gs_reference.do_move((5, y), go.WHITE)
# board configuration and liberties of gs_capture and of gs_reference should be identical
self.assertTrue(np.all(gs_reference.board == gs_capture.board))
self.assertTrue(
np.all(gs_reference.liberty_counts == gs_capture.liberty_counts))
def setUp(self):
self.gs = GameState()
self.mcts = MCTS(dummy_value, dummy_policy, dummy_rollout, n_playout=2)
def test_true_eye(self):
gs = GameState(size=7)
gs.do_move((1, 0), go.BLACK)
gs.do_move((0, 1), go.BLACK)
# false eye at 0, 0
self.assertTrue(gs.is_eyeish((0, 0), go.BLACK))
self.assertFalse(gs.is_eye((0, 0), go.BLACK))
# make it a true eye by turning the corner (1, 1) into an eye itself
gs.do_move((1, 2), go.BLACK)
gs.do_move((2, 1), go.BLACK)
gs.do_move((2, 2), go.BLACK)
gs.do_move((0, 2), go.BLACK)
self.assertTrue(gs.is_eyeish((0, 0), go.BLACK))
self.assertTrue(gs.is_eye((0, 0), go.BLACK))
self.assertTrue(gs.is_eye((1, 1), go.BLACK))
def test_standard_ko(self):
# . B . .
# B X B .
# W B W .
# . W . .
gs = GameState(size=9)
gs.do_move((1, 0)) # B
gs.do_move((2, 0)) # W
gs.do_move((2, 1)) # B
gs.do_move((3, 1)) # W
gs.do_move((1, 2)) # B
gs.do_move((2, 2)) # W
gs.do_move((0, 1)) # B
gs.do_move((1, 1)) # W trigger capture and ko
self.assertEqual(gs.num_black_prisoners, 1)
self.assertEqual(gs.num_white_prisoners, 0)
self.assertFalse(gs.is_legal((2, 1)))
gs.do_move((5, 5))
gs.do_move((5, 6))
self.assertTrue(gs.is_legal((2, 1)))
def test_batch_eval_state(self):
policy = ResnetPolicy(["board", "liberties", "sensibleness", "capture_size"])
results = policy.batch_eval_state([GameState(), GameState()])
self.assertEqual(len(results), 2) # one result per GameState
self.assertEqual(len(results[0]), 361) # each one has 361 (move,prob) pairs
def test_default_policy(self):
policy = ResnetPolicy(["board", "liberties", "sensibleness", "capture_size"])
policy.eval_state(GameState())
def setUp(self):
self.gs = GameState()
self.node = TreeNode(None, 1.0)
class TestMCTS(unittest.TestCase):
def setUp(self):
self.gs = GameState()
self.mcts = MCTS(dummy_value, dummy_policy, dummy_rollout, n_playout=2)
def _count_expansions(self):
"""Helper function to count the number of expansions past the root using the dummy policy
"""
node = self.mcts._root
expansions = 0
# Loop over actions in decreasing probability.
for action, _ in sorted(dummy_policy(self.gs),
key=lambda (a, p): p,
reverse=True):
if action in node._children:
expansions += 1
node = node._children[action]
else:
break
return expansions
def test_playout(self):
self.mcts._playout(self.gs.copy(), 8)
# Assert that the most likely child was visited (according to the dummy policy below).
self.assertEqual(1, self.mcts._root._children[(18, 18)]._n_visits)
# Assert that the search depth expanded nodes 8 times.
self.assertEqual(8, self._count_expansions())
def test_playout_with_pass(self):
# Test that playout handles the end of the game (i.e. passing/no moves). Mock this by
# creating a policy that returns nothing after 4 moves.
def stop_early_policy(state):
if len(state.history) <= 4:
return dummy_policy(state)
else:
return []
self.mcts = MCTS(dummy_value,
stop_early_policy,
stop_early_policy,
n_playout=2)
self.mcts._playout(self.gs.copy(), 8)
# Assert that (18, 18) and (18, 17) are still only visited once.
self.assertEqual(1, self.mcts._root._children[(18, 18)]._n_visits)
# Assert that no expansions happened after reaching the "end" in 4 moves.
self.assertEqual(5, self._count_expansions())
def test_get_move(self):
move = self.mcts.get_move(self.gs)
self.mcts.update_with_move(move)
# success if no errors
def test_update_with_move(self):
move = self.mcts.get_move(self.gs)
self.gs.do_move(move)
self.mcts.update_with_move(move)
# Assert that the new root still has children.
self.assertTrue(len(self.mcts._root._children) > 0)
# Assert that the new root has no parent (the rest of the tree will be garbage collected).
self.assertIsNone(self.mcts._root._parent)
# Assert that the next best move according to the root is (18, 17), according to the
# dummy policy below.
self.assertEqual((18, 17), self.mcts._root.select()[0])
class TestLiberties(unittest.TestCase):
def setUp(self):
# 0 1 2 3 4 5 6 7 8 9 A B
# 0 . . . . . . . . . . . .
# 1 . . . . . . . . . . . .
# 2 . . . . . . . . . . . .
# 3 . . . . . . . . . . . .
# 4 . . . . . B B . . . . .
# 5 . . . . . W B . . . . .
# 6 . . . . . . B . . . . .
# 7 . . . . . . . . . . . .
# 8 . . . . . . . . . . . .
# 9 . . . . . . . . . . W .
# A . . . . . . . . . . W W
# B . . . . . . . . . . . .
self.s = GameState()
self.s.do_move((4, 5))
self.s.do_move((5, 5))
self.s.do_move((5, 6))
self.s.do_move((10, 10))
self.s.do_move((4, 6))
self.s.do_move((10, 11))
self.s.do_move((6, 6))
self.s.do_move((9, 10))
def test_curr_liberties(self):
self.assertEqual(self.s.liberty_counts[5][5], 2)
self.assertEqual(self.s.liberty_counts[4][5], 8)
self.assertEqual(self.s.liberty_counts[5][6], 8)
def test_neighbors_edge_cases(self):
st = GameState()
st.do_move((0, 0)) # B B . . . . .
st.do_move((5, 5)) # B W . . . . .
st.do_move((0, 1)) # . . . . . . .
st.do_move((6, 6)) # . . . . . . .
st.do_move((1, 0)) # . . . . . W .
st.do_move((1, 1)) # . . . . . . W
# get_group in the corner
self.assertEqual(len(st.get_group((0, 0))), 3, "group size in corner")
# get_group of an empty space
self.assertEqual(len(st.get_group((4, 4))), 0,
"group size of empty space")
# get_group of a single piece
self.assertEqual(len(st.get_group((5, 5))), 1,
"group size of single piece")
def test_neighbors_edge_cases(self):
st = GameState()
st.do_move((0, 0)) # B B . . . . .
st.do_move((5, 5)) # B W . . . . .
st.do_move((0, 1)) # . . . . . . .
st.do_move((6, 6)) # . . . . . . .
st.do_move((1, 0)) # . . . . . W .
st.do_move((1, 1)) # . . . . . . W
# get_group in the corner
self.assertEqual(len(st.get_group((0, 0))), 3, "group size in corner")
# get_group of an empty space
self.assertEqual(len(st.get_group((4, 4))), 0,
"group size of empty space")
# get_group of a single piece
self.assertEqual(len(st.get_group((5, 5))), 1,
"group size of single piece")
def play_batch(player_RL, player_SL, batch_size, features):
"""Play a batch of games in parallel and return one training pair
from each game.
"""
def do_move(states, moves):
for st, mv in zip(states, moves):
if not st.is_end_of_game:
# Only do more moves if not end of game already
st.do_move(mv)
return states
def do_rand_move(states, player, player_RL):
"""Do a uniform-random move over legal moves and record info for
training. Only gets called once per game.
"""
colors = [st.current_player for st in states] # Record player color
legal_moves = [st.get_legal_moves() for st in states]
rand_moves = [lm[np.random.choice(len(lm))] for lm in legal_moves]
states = do_move(states, rand_moves)
player = player_RL
X_list = [st.copy() for st in states] # For later 1hot preprocessing
return X_list, colors, states, player
def convert(X_list, preprocessor):
"""Convert states to 1-hot and concatenate. X's are game state objects.
"""
states = np.concatenate(
[preprocessor.state_to_tensor(X) for X in X_list], axis=0)
return states
# Lists of game training pairs (1-hot)
preprocessor = Preprocess(features)
player = player_SL
states = [GameState() for i in xrange(batch_size)]
# Randomly choose turn to play uniform random. Move prior will be from SL
# policy. Moves after will be from RL policy.
i_rand_move = np.random.choice(range(450))
X_list = None
winners = None
turn = 0
while True:
# Do moves (black)
if turn == i_rand_move:
# Make random move, then switch from SL to RL policy
X_list, colors, states, player = do_rand_move(states, player,
player_RL)
else:
# Get moves (batch)
moves_black = player.get_moves(states)
# Do moves (black)
states = do_move(states, moves_black)
turn += 1
# Do moves (white)
if turn == i_rand_move:
# Make random move, then switch from SL to RL policy
X_list, colors, states, player = do_rand_move(states, player,
player_RL)
else:
moves_white = player.get_moves(states)
states = do_move(states, moves_white)
turn += 1
# If all games have ended, we're done. Get winners.
done = [st.is_end_of_game or st.turns_played > 500 for st in states]
print turn
if all(done):
break
# Concatenate training examples
X = None
if X_list is not None:
X = convert(X_list, preprocessor)
winners = np.array([st.get_winner() for st in states]).reshape(batch_size, 1)
return X, winners
def test_simple_eye(self):
# create a black eye in top left (1, 1), white in bottom right (5, 5)
gs = GameState(size=7)
gs.do_move((1, 0)) # B
gs.do_move((5, 4)) # W
gs.do_move((2, 1)) # B
gs.do_move((6, 5)) # W
gs.do_move((1, 2)) # B
gs.do_move((5, 6)) # W
gs.do_move((0, 1)) # B
gs.do_move((4, 5)) # W
# test black eye top left
self.assertTrue(gs.is_eyeish((1, 1), go.BLACK))
self.assertFalse(gs.is_eyeish((1, 1), go.WHITE))
# test white eye bottom right
self.assertTrue(gs.is_eyeish((5, 5), go.WHITE))
self.assertFalse(gs.is_eyeish((5, 5), go.BLACK))
# test no eye in other random positions
self.assertFalse(gs.is_eyeish((1, 0), go.BLACK))
self.assertFalse(gs.is_eyeish((1, 0), go.WHITE))
self.assertFalse(gs.is_eyeish((2, 2), go.BLACK))
self.assertFalse(gs.is_eyeish((2, 2), go.WHITE))
def run_n_games(optimizer, learner, opponent, num_games):
'''Run num_games games to completion, calling train_batch() on each position
the learner sees.
(Note: optimizer only accumulates gradients in its update function until
all games have finished)
'''
board_size = learner.policy.model.input_shape[-1]
states = [GameState(size=board_size) for _ in range(num_games)]
learner_net = learner.policy.model
# Start all odd games with moves by 'opponent'. Even games will have 'learner' black.
learner_color = [
go.BLACK if i % 2 == 0 else go.WHITE for i in range(num_games)
]
odd_states = states[1::2]
moves = opponent.get_moves(odd_states)
for st, mv in zip(odd_states, moves):
st.do_move(mv)
current = learner
other = opponent
# Need to keep track of the index of unfinished states so that we can communicate which one is
# being updated to the optimizer.
idxs_to_unfinished_states = {i: states[i] for i in range(num_games)}
while len(idxs_to_unfinished_states) > 0:
# Get next moves by current player for all unfinished states.
moves = current.get_moves(idxs_to_unfinished_states.values())
just_finished = []
# Do each move to each state in order.
for (idx, state), mv in zip(idxs_to_unfinished_states.iteritems(),
moves):
# Order is important here. We must first get the training pair on the unmodified state.
# Next, the state is updated and checked to see if the game is over. If it is over, the
# optimizer is notified via set_result. Finally, train_on_batch is called, which
# will trigger an update of all parameters only if set_result() has been called
# for all games already (so set_result must come before train_on_batch).
is_learnable = current is learner and mv is not go.PASS_MOVE
if is_learnable:
(X, y) = _make_training_pair(state, mv,
learner.policy.preprocessor)
state.do_move(mv)
if state.is_end_of_game:
learner_is_winner = state.get_winner() == learner_color[idx]
optimizer.set_result(idx, learner_is_winner)
just_finished.append(idx)
if is_learnable:
optimizer.set_current_game(idx)
learner_net.train_on_batch(X, y)
# Remove games that have finished from dict.
for idx in just_finished:
del idxs_to_unfinished_states[idx]
# Swap 'current' and 'other' for next turn.
current, other = other, current
# Return the win ratio.
wins = sum(state.get_winner() == pc
for (state, pc) in zip(states, learner_color))
return float(wins) / num_games