def produce_data(when_to_start, dump_p, n_games, filename, folder):
list_of_agents = [RandomAgent(), GreedyAgentBoost(), MinMaxAgent()]
arek = ArenaMultiThread()
arek.start_collecting_states()
arek.collect_only_from_middle_game(when_to_start, dump_p)
arek.all_vs_all('deterministic', list_of_agents, n_games)
arek.dump_collected_states(filename, folder)
def run_baseline_comparison_v1():
experiment_name = 'baseline_comparison_v1'
agent1 = GreedyAgentBoost(weight=[100, 1.5, 2.5, 1, 0.1])
agent2 = GreedyAgentBoost(
weight=[0.99953495, 0.02010871, 0.02010487, 0.01095619, 0.00113329])
agent3 = MinMaxAgent(weight=[100, 2, 2, 1, 0.1])
# agent4 = MinMaxAgent(weight=[100, 2, 2, 1, 0.1], decay=0.7)
# agent5 = GreedySearchAgent(depth = 9, weight = [100,2,2,1,0.1], decay = 0.95)
# agent6 = GreedySearchAgent(depth = 7, weight = [100,2,2,1,0.1])
agent7 = GreedySearchAgent(
depth=7,
weight=[0.99953495, 0.02010871, 0.02010487, 0.01095619, 0.00113329],
decay=0.95)
# agent8 = RandomAgent(distribution='first_buy')
multi_arena = ArenaMultiThread()
n_games = 1
list_of_agents = [agent1, agent2, agent3, agent7]
results = multi_arena.all_vs_all(list_of_agents, n_games)
if main_thread:
print(' \n \n {}'.format(results.to_pandas()))
print('\n \n \n')
print(results)
wins = results.to_pandas(param='wins').to_csv('wins.csv')
vic_points = results.to_pandas(
param='victory_points').to_csv('victory_points.csv')
rewards = results.to_pandas(param='reward').to_csv('reward.csv')
#leader_board = LeaderBoard(list_of_agents)
#leader_board.load_from_file()
#leader_board.register_from_games_statistics(results)
#print(leader_board)
#leader_board.save_to_file()
plt.title('Average win rate over {} games per pair:'.format(2 *
n_games))
wins_pic = results.create_heatmap(param='wins', average=True)
plt.savefig('reports/wins.png')
plt.clf()
plt.title('Average reward over {} games per pair:'.format(2 * n_games))
reward_pic = results.create_heatmap('reward', average=True)
plt.savefig('reports/reward.png')
plt.clf()
plt.title('Average victory points over {} games per pair:'.format(
2 * n_games))
vic_points_pic = results.create_heatmap('victory_points', average=True)
plt.savefig('reports/victory_points.png')
plt.clf()
def go():
arena = MultiArena()
data_transformer = IdentityTransformer()
model = StateEncoder(final_layer=ValueRegressor(),
data_transformer=data_transformer)
model.load_weights('archive/weights_tt1/epoch_41.h5')
value_policy = ValueEvaluator(model=model, weights_file=None)
mcts_agent = SingleMCTSAgent(50,
value_policy,
0.41,
create_visualizer=False,
show_unvisited_nodes=False,
log_to_neptune=False)
opp = MinMaxAgent()
results = arena.run_many_duels('deterministic', [mcts_agent, opp], 10, 1,
True)
print(results)
def __init__(self,
gems_encoder_dim : int = None,
price_encoder_dim : int = None,
profit_encoder_dim : int = None,
cards_points_dim: int = None,
cards_dense1_dim: int = None,
cards_dense2_dim: int = None,
board_nobles_dense1_dim : int = None,
board_nobles_dense2_dim : int = None,
full_board_dense1_dim: int = None,
full_board_dense2_dim: int = None,
player_points_dim: int = None,
player_nobles_dim: int = None,
full_player_dense1_dim: int = None,
full_player_dense2_dim: int = None,
final_layer= None,
data_transformer = None,
network_name: str = None
):
super().__init__()
self.vectorizer = Vectorizer()
self.final_layer = final_layer
self.data_transformer = data_transformer
self.params['data transormation'] = self.data_transformer.name
self.params['final layer name'] = self.final_layer.name
self.params['gems_encoder_dim'] = gems_encoder_dim
self.params['gems_encoder_dim'] = gems_encoder_dim
self.params['price_encoder_dim'] = price_encoder_dim
self.params['profit_encoder_dim'] = profit_encoder_dim
self.params['cards_points_dim'] = cards_points_dim
self.params['cards_dense1_dim'] = cards_dense1_dim
self.params['cards_dense2_dim'] = cards_dense2_dim
self.params['board_nobles_dense1_dim'] = board_nobles_dense1_dim
self.params['board_nobles_dense2_dim'] = board_nobles_dense2_dim
self.params['full_board_dense1_dim']= full_board_dense1_dim
self.params['full_board_dense2_dim'] = full_board_dense2_dim
self.params['player_points_dim'] = player_points_dim
self.params['player_nobles_dim'] = player_nobles_dim
self.params['full_player_dense1_dim'] = full_player_dense1_dim
self.params['full_player_dense2_dim']= full_player_dense2_dim
self.arena = Arena()
self.network_agent = ValueNNAgent(self)
self.easy_opp = RandomAgent(distribution='first_buy')
self.medium_opp = GreedyAgentBoost()
self.hard_opp = MinMaxAgent()
self.neptune_monitor = NeptuneMonitor()
self.network_name = network_name
self.gems_encoder = GemsEncoder(gems_encoder_dim)
self.price_encoder = PriceEncoder(price_encoder_dim)
self.board_encoder = BoardEncoder(self.gems_encoder,
ManyNoblesEncoder(price_encoder_dim,
board_nobles_dense1_dim,
board_nobles_dense2_dim),
ManyCardsEncoder(MAX_CARDS_ON_BORD,
profit_encoder_dim,
price_encoder_dim,
cards_points_dim,
cards_dense1_dim,
cards_dense2_dim
),
full_board_dense1_dim,
full_board_dense2_dim)
self.player_encoder = PlayerEncoder(self.gems_encoder,
self.price_encoder,
ManyCardsEncoder(MAX_RESERVED_CARDS,
profit_encoder_dim,
price_encoder_dim,
cards_points_dim,
cards_dense1_dim,
cards_dense2_dim
),
player_points_dim,
player_nobles_dim,
full_player_dense1_dim,
full_player_dense2_dim)
active_player_input = PlayersInputGenerator('active_').inputs
other_player_input = PlayersInputGenerator('other_').inputs
board_input = self.board_encoder.inputs
self.inputs = board_input + active_player_input + other_player_input
board_encoded = self.board_encoder(board_input)
active_player_encoded = self.player_encoder(active_player_input)
other_player_encoded = self.player_encoder(other_player_input)
full_state = Concatenate(axis=-1)([board_encoded, active_player_encoded, other_player_encoded])
full_state = Dense(full_player_dense1_dim, activation='relu')(full_state)
final_state = Dense(full_player_dense2_dim, activation='relu')(full_state)
result = self.final_layer(final_state)
self.layer = Model(inputs = self.inputs, outputs = final_state, name = 'full_state_splendor_estimator')
self.network = Model(inputs = self.inputs, outputs = result, name = 'full_state_splendor_estimator')
self.network.compile(Adam(), loss='mean_squared_error')
self.params['Model name'] = 'Average pooling model'
self.params['optimizer_name'] = 'Adam'
from agents.random_agent import RandomAgent
from agents.minmax_agent import MinMaxAgent
from agents.greedysearch_agent2 import GreedySearchAgent
from arena.arena import Arena
environment_id = 'gym_splendor_code:splendor-v1'
fight_pit = Arena(environment_id)
goku = RandomAgent(distribution='first_buy')
goku2 = RandomAgent(distribution='uniform')
#gohan = RandomAgent(distribution='uniform_on_types')
#gohan = RandomAgent(distribution='uniform')
#goku = GreedyAgen/t(weight = 0.3)
gohan = GreedySearchAgent(depth=5)
goku = MinMaxAgent(name="MinMax", depth=3)
gohan.name = "g2"
goku.name = "g1"
# profi = cProfile.Profile()
#
# profi.run('(fight_pit.run_many_duels([goku, gohan], number_of_games=50))')
# profi.dump_stats('profi2.prof')
fight_pit.run_one_duel([goku, gohan], render_game=True)
# time_dupa = time.time()
# for i in range(100):
# print(i)
# fight_pit = Arena()
# fight_pit.run_one_duel([goku, gohan], starting_agent_id=0)
# print(time.time() - time_dupa)
gohan = GreedyAgentBoost(weight = [100,2.5,1.5,1,0.1]) print(gohan.name) goku = RandomAgent(distribution='uniform') print(fight_pit.run_many_duels([goku, gohan], number_of_games = n_games, shuffle_agents=True)) goku = RandomAgent(distribution='uniform_on_types') print(fight_pit.run_many_duels([goku, gohan], number_of_games = n_games, shuffle_agents=True)) goku = RandomAgent(distribution = 'first_buy') print(fight_pit.run_many_duels([goku, gohan], number_of_games = n_games, shuffle_agents=True)) gohan = MinMaxAgent(name = "MinMax", depth = 2) print(gohan.name) goku = RandomAgent(distribution='uniform') print(fight_pit.run_many_duels([goku, gohan], number_of_games = n_games, shuffle_agents=True)) goku = RandomAgent(distribution='uniform_on_types') print(fight_pit.run_many_duels([goku, gohan], number_of_games = n_games, shuffle_agents=True)) goku = RandomAgent(distribution = 'first_buy') print(fight_pit.run_many_duels([goku, gohan], number_of_games = n_games, shuffle_agents=True)) gohan = MinMaxAgent(name = "MinMax", depth = 3) print(gohan.name)