def test_features(self):
players = [
LearningPlayer(name='random',
estimation_mode=LearningPlayer.ACTUAL_Q)
for _ in range(3)
]
game = LandlordGame(players=players)
while not game.is_round_over():
curr_player = game.get_current_player()
curr_features = curr_player._derive_features(game)
curr_hand_vector = game.get_current_player().get_hand_vector(
game, game.get_current_position())
move = game.get_current_player().make_move(
game, game.get_current_position())
curr_move_vector = game.get_current_player().compute_move_vector(
game.get_current_position(), game.get_landlord_position(),
move)
game.play_move(move)
self.assertTrue(
np.allclose(curr_features,
curr_player.record_history_matrices[-1]))
self.assertTrue(
np.allclose(curr_move_vector,
curr_player.record_move_vectors[-1]))
self.assertTrue(
np.allclose(curr_hand_vector,
curr_player.record_hand_vectors[-1]))
def test_features_v2(self):
players = [
LearningPlayer_v2(name='random',
epsilon=0,
estimation_mode=LearningPlayer.ACTUAL_Q,
learning_rate=1) for _ in range(3)
]
game = LandlordGame(players=players)
while not game.is_round_over():
curr_player = game.get_current_player()
curr_features = curr_player._derive_features(game)
best_move, best_move_q = curr_player.decide_best_move(game)
curr_move_vector = game.get_current_player().compute_move_vector(
game.get_current_position(), game.get_landlord_position(),
best_move)
curr_hand_vector = game.get_current_player(
).compute_remaining_hand_vector(game, curr_move_vector,
game.get_current_position())
curr_player.record_move(game, best_move, best_move_q,
game.get_current_position())
game.play_move(best_move)
self.assertTrue(
np.allclose(curr_features,
curr_player.record_history_matrices[-1]))
self.assertTrue(
np.allclose(curr_move_vector,
curr_player.record_move_vectors[-1]))
self.assertTrue(
np.allclose(curr_hand_vector,
curr_player.record_hand_vectors[-1]))
players[0].compute_future_q(game)
if game.has_winners():
print(np.max(np.abs(players[0].get_estimated_qs())))
self.assertTrue(np.max(np.abs(players[0].get_estimated_qs())) == 1)
self.assertTrue(
players[0].record_history_matrices[0][0].dtype == np.int8)
def test_record_actual_q(self):
def load_best_sim_net(net):
return LearningPlayer(name=net,
net_dir='../models/' + net,
estimation_mode=LearningPlayer.ACTUAL_Q,
epsilon=0,
discount_factor=1)
players = [load_best_sim_net('4_2_sim4_model10') for i in range(3)]
player_0_scores = []
game = LandlordGame(players=players)
while not game.is_round_over():
curr_player = game.get_current_player()
best_move, best_move_q = curr_player.decide_best_move(game)
if curr_player == players[0]:
player_0_scores.append(best_move_q)
curr_player.make_move(game)
game.play_move(best_move)
for player in players:
player.compute_future_q(game)
record_state = players[0]._record_state_q
future_q = players[0].get_estimated_qs()
# assert in bounds based on update function
for i, val in enumerate(record_state):
if i != len(record_state) - 1:
if record_state[i + 1] < record_state[i]:
self.assertTrue(
record_state[i + 1] < future_q[i] < record_state[i])
elif record_state[i + 1] > record_state[i]:
self.assertTrue(
record_state[i + 1] > future_q[i] > record_state[i])
self.assertEqual(len(players[0].get_record_hand_vectors()),
len(players[0].get_estimated_qs()))
def human_game(player_names, perspective):
perspective_hand = None
players = []
for player_name in player_names:
if player_name == perspective:
perspective_hand = manual_hand()
player_is_perspective = (player_name == perspective)
players.append(
HumanPlayer(name=player_name,
reference_player=reference_player,
known_hand=player_is_perspective,
ai_before=player_is_perspective))
game = LandlordGame(players, kitty_callback=manual_kitty)
first_player = get_first_player(game)
game.force_current_position(first_player)
game.force_hand(perspective_position(game, perspective), perspective_hand)
while not game.is_round_over():
current_player = game.get_current_player()
current_position = game.get_current_position()
best_move, best_move_q = current_player.decide_best_move(game)
print(current_player.get_name(), "(" + game.get_position_role_name(current_position) + ", " \
+ str(len(game.get_hand(current_position))) + "):", best_move, '(' + str(best_move_q) + ')')
# play with known hand if it matches perspective
game.play_move(best_move,
hand_known=current_player.get_name() == perspective)
if type(game.get_last_played()) == KittyReveal:
print(game.get_last_played())
if game.has_winners():
for winner in game.get_winners():
print('WINNERS:', game.get_ai_players()[winner].get_name())
def play_against_two(players, show_q=True):
game = LandlordGame(players)
while not game.is_round_over():
current_player = game.get_current_player()
current_position = game.get_current_position()
best_move, best_move_q = current_player.decide_best_move(game)
if show_q:
best_move_q_str = '(' + str(best_move_q) + ')'
else:
best_move_q_str = ''
print(current_player.get_name(), "(" + game.get_position_role_name(current_position) + ", " \
+ str(len(game.get_hand(current_position))) + "):", best_move, best_move_q_str)
game.play_move(best_move)
if type(game.get_last_played()) == KittyReveal:
print(game.get_last_played())
if game.has_winners():
for winner in game.get_winners():
print('WINNERS:', game.get_ai_players()[winner].get_name())
def test_self_feed(self):
players = [self.load_v2_net("4_8_actualq1_model20") for _ in range(3)]
#players = [self.load_v2_net("4_2_sim4_model15") for _ in range(3)]
game = LandlordGame(players=players)
best_move_qs = []
all_history_features = []
history_vectors = []
all_hand_vectors = []
all_move_vectors = []
while not game.is_round_over():
best_move, best_move_q = game.get_current_player(
).decide_best_move(game, game.get_current_position())
game.get_current_player().record_move(game, best_move, best_move_q,
game.get_current_position())
if game.get_current_player() == players[0]:
history_features = players[0]._derive_features(game)
all_history_features.append(history_features)
# all the moves we make from here will not affect the history, so assess it and copy
history_vectors.append(players[0].history_net.predict(
np.array([history_features]), batch_size=1)[0])
# create features for each of the possible moves from this position
all_move_vectors.append(players[0].compute_move_vector(
game.get_current_position(), game.get_landlord_position(),
best_move))
all_hand_vectors.append(
players[0].compute_remaining_hand_vector(
game, all_move_vectors[-1],
game.get_current_position()))
predicted_q = players[0].position_net.predict([
np.array([history_vectors[-1]]),
np.array([all_move_vectors[-1]]),
np.array([all_hand_vectors[-1]])
])[0][0]
self.assertAlmostEqual(predicted_q, best_move_q, places=4)
best_move_qs.append(best_move_q)
game.play_move(best_move)
players[0].compute_future_q(game)
history_matrices = players[0].get_record_history_matrices()
for i, j in zip(all_history_features, history_matrices):
self.assertTrue(np.allclose(i, j))
move_vectors = players[0].get_record_move_vectors()
for i, j in zip(all_move_vectors, move_vectors):
self.assertTrue(np.allclose(i, j))
hand_vectors = players[0].get_record_hand_vectors()
for i, j in zip(all_hand_vectors, hand_vectors):
self.assertTrue(np.allclose(i, j))
qs = players[0].get_estimated_qs()
pred_qs = []
# recreate
for i, records in enumerate(
zip(history_matrices, move_vectors, hand_vectors, qs)):
history_matrix, move_vector, hand_vector, q = records
history_vector = players[0].history_net.predict(
np.array([history_matrix]))[0]
self.assertTrue(np.allclose(history_vector, history_vectors[i]))
pred_qs.append(players[0].position_net.predict([[history_vector],
[move_vector],
[hand_vector]
])[0][0])
# works only if learning rate is 0
self.assertTrue(np.allclose(qs, pred_qs))