def main():
game_state = sample_game_states[18]
num_simulations = 100
ucb_constant = 0.1
player = UCTPlayer(ucb_const=ucb_constant, num_samples=1, num_simulations=num_simulations)
root_node = MCNode(game_state=game_state)
mc_tree = MCTree(root_node=root_node)
for sim_num in range(1, player.num_simulations + 1):
selected_node = player.selection(mc_tree)
rewards = player.simulation(selected_node)
mc_tree.backup_rewards(leaf_node=selected_node, rewards=rewards)
mc_tree.visualize_tree(ucb=player.ucb_const,
filename="img_{}".format(sim_num))
filenames = ["img_{}.png".format(i) for i in range(player.num_simulations, 0, -1)]
gif_name = "Gif_test.gif"
duration = 0.2
make_gif(filenames=filenames, outputname=gif_name, duration=duration)
im = Image.open(gif_name)
frames = [frame.copy() for frame in ImageSequence.Iterator(im)]
frames.reverse()
frames[0].save('reversed.gif', save_all=True, append_images=frames[1:])
def test_add_node(game_state_partner, next_state, different_next_state):
root_node = MCNode(game_state=game_state_partner, parent=None, previous_action= None)
child = MCNode(game_state=next_state, parent=root_node, previous_action=(NO_GAME, None))
root_node.add_child(child)
assert len(root_node.children) == 1
child.update_rewards([30, 30, -30, -30])
child.update_visits()
root_node.update_visits()
assert child.current_player == 1
assert child.get_average_reward(child.current_player) == 30
child.update_rewards([20, 20, -20, -20])
child.update_visits()
root_node.update_visits()
assert child.get_average_reward(child.current_player) == 25
sec_child = MCNode(game_state=different_next_state, parent=root_node, previous_action=(PARTNER_MODE, LEAVES))
assert not root_node.fully_expanded()
sec_child.update_rewards([-20, 20, 20, -20])
sec_child.update_visits()
root_node.update_visits()
assert root_node.best_child(ucb_const=1).previous_action == (NO_GAME, None)
def expand(self, mc_tree, node):
not_visited_actions = set(node.game_state["possible_actions"])
for child in node.children:
not_visited_actions.remove(child.previous_action)
chosen_action = random.choice(tuple(not_visited_actions))
new_state = self.get_new_state(game_state=node.game_state,
action=chosen_action)
new_node = MCNode(parent=node, game_state=new_state, previous_action=chosen_action)
mc_tree.add_node(node=new_node,
parent_node=node)
return new_node
def uct_search(self, game_state):
root_node = MCNode(game_state=game_state)
mc_tree = MCTree(root_node=root_node)
for sim_num in range(1, self.num_simulations + 1):
selected_node = self.selection(mc_tree)
rewards = self.simulation(selected_node)
mc_tree.backup_rewards(leaf_node=selected_node, rewards=rewards)
results = []
for child in mc_tree.root_node.children:
results.append((child.previous_action, child.visits, child.get_average_reward(root_node.current_player)))
return results
def main():
sim_player_list = [
NNPlayer(
game_mode_nn='../players/models/bigger_classifier200.hdf5',
partner_nn='../players/models/partner_model_wider_data_2.hdf5',
solo_nn='../players/models/solo_model_wider_data_10.hdf5',
wenz_nn='../players/models/wenz_model_wider_data_10.hdf5'),
NNPlayer(
game_mode_nn='../players/models/bigger_classifier200.hdf5',
partner_nn='../players/models/partner_model_wider_data_2.hdf5',
solo_nn='../players/models/solo_model_wider_data_10.hdf5',
wenz_nn='../players/models/wenz_model_wider_data_10.hdf5'),
NNPlayer(
game_mode_nn='../players/models/bigger_classifier200.hdf5',
partner_nn='../players/models/partner_model_wider_data_2.hdf5',
solo_nn='../players/models/solo_model_wider_data_10.hdf5',
wenz_nn='../players/models/wenz_model_wider_data_10.hdf5'),
NNPlayer(
game_mode_nn='../players/models/bigger_classifier200.hdf5',
partner_nn='../players/models/partner_model_wider_data_2.hdf5',
solo_nn='../players/models/solo_model_wider_data_10.hdf5',
wenz_nn='../players/models/wenz_model_wider_data_10.hdf5')
]
game_state = sample_game_states[15]
num_simulations = 100
ucb_constant = 0.1
player = UCTPlayer(ucb_const=ucb_constant,
num_samples=1,
num_simulations=num_simulations,
simulation_player_list=None)
root_node = MCNode(game_state=game_state)
mc_tree = MCTree(root_node=root_node)
for sim_num in range(1, player.num_simulations + 1):
selected_node = player.selection(mc_tree)
rewards = player.simulation(selected_node)
mc_tree.backup_rewards(leaf_node=selected_node, rewards=rewards)
mc_tree.visualize_tree(
ucb=player.ucb_const,
filename="Tree_{}nodes{}ucb_const{}game_mode".format(
num_simulations, ucb_constant, game_state["game_mode"]))
def main():
sim_player_list = [NNPlayer(game_mode_nn='../players/models/bigger_classifier200.hdf5',
partner_nn='../players/models/partner_model_wider_data_2.hdf5',
solo_nn='../players/models/solo_model_wider_data_10.hdf5',
wenz_nn='../players/models/wenz_model_wider_data_10.hdf5'),
NNPlayer(game_mode_nn='../players/models/bigger_classifier200.hdf5',
partner_nn='../players/models/partner_model_wider_data_2.hdf5',
solo_nn='../players/models/solo_model_wider_data_10.hdf5',
wenz_nn='../players/models/wenz_model_wider_data_10.hdf5'),
NNPlayer(game_mode_nn='../players/models/bigger_classifier200.hdf5',
partner_nn='../players/models/partner_model_wider_data_2.hdf5',
solo_nn='../players/models/solo_model_wider_data_10.hdf5',
wenz_nn='../players/models/wenz_model_wider_data_10.hdf5'),
NNPlayer(game_mode_nn='../players/models/bigger_classifier200.hdf5',
partner_nn='../players/models/partner_model_wider_data_2.hdf5',
solo_nn='../players/models/solo_model_wider_data_10.hdf5',
wenz_nn='../players/models/wenz_model_wider_data_10.hdf5')]
cum_depth = 0
for game_state in sample_game_states:
num_simulations = 100
ucb_constant = 0.1
player = UCTPlayer(ucb_const=ucb_constant,
num_samples=1,
num_simulations=num_simulations,
simulation_player_list=None)
root_node = MCNode(game_state=game_state)
mc_tree = MCTree(root_node=root_node)
for sim_num in range(1, player.num_simulations + 1):
selected_node = player.selection(mc_tree)
rewards = player.simulation(selected_node)
mc_tree.backup_rewards(leaf_node=selected_node, rewards=rewards)
depth = mc_tree.max_depth()
cum_depth += depth
print('Average tree depth without NN:', cum_depth / len(sample_game_states))
cum_depth_with_nn = 0
for game_state in sample_game_states:
num_simulations = 100
ucb_constant = 0.1
player = UCTPlayer(ucb_const=ucb_constant,
num_samples=1,
num_simulations=num_simulations,
simulation_player_list=sim_player_list)
root_node = MCNode(game_state=game_state)
mc_tree = MCTree(root_node=root_node)
for sim_num in range(1, player.num_simulations + 1):
selected_node = player.selection(mc_tree)
rewards = player.simulation(selected_node)
mc_tree.backup_rewards(leaf_node=selected_node, rewards=rewards)
depth = mc_tree.max_depth()
cum_depth_with_nn += depth
print('Average tree depth with NN:', cum_depth_with_nn / len(sample_game_states))