for timestep in range(timesteps): action = agent.step(state) next_state, reward, done = env.step(action) ts = (state, action, reward, next_state, done) agent.feed(ts) if timestep % evaluate_every == 0: rewards = [] state = eval_env.reset() for _ in range(evaluate_num): action, _ = agent.eval_step(state) _, reward, done = env.step(action) if done: rewards.append(reward) logger.log_performance(env.timestep, np.mean(rewards)) # Close files in the logger logger.close_files() # Plot the learning curve logger.plot('DQN') # Save model save_dir = 'models/leduc_holdem_single_dqn' if not os.path.exists(save_dir): os.makedirs(save_dir) saver = tf.compat.v1.train.Saver() saver.save(sess, os.path.join(save_dir, 'model'))
for episode in range(episode_num): print("Episode: " + str(episode)) # First sample a policy for the episode for agent in agents: agent.sample_episode_policy() # Generate data from the environment trajectories, _ = env.run(is_training=True) # Feed transitions into agent memory, and train the agent for i in range(env.player_num): for ts in trajectories[i]: agents[i].feed(ts) # Evaluate the performance. Play with random agents. if episode % evaluate_every == 0: a, b = env.timestep, tournament(eval_env, evaluate_num)[0] logger.log_performance(a, b) csvw.writerow([a, b]) f.flush() # saver.save(sess, os.path.join(save_dir, 'model_' + str(episode))) # logger.log_performance(env.timestep, tournament(eval_env, evaluate_num)[0]) # Close files in the logger logger.close_files() # Plot the learning curve logger.plot('NFSP')
# Feed transitions into agent memory, and train the agent for ts in trajectories[0]: agent.feed(ts) # Evaluate the performance. Play with random agents. if episode % evaluate_every == 0: if episode > 0: current_time = time.time() episodes_per_sec = episode / (current_time - start_time) remaining_mins = (episode_num - episode) / episodes_per_sec / 60 print( f"Current Rate: {episodes_per_sec:.2f}, Estimated Time Remaining: {remaining_mins:.2f} mins" ) reward = tournament(eval_env, evaluate_num)[0] logger.log_performance(env.timestep, reward) with open(os.path.join(log_dir, "perf.csv"), "a+") as fd: fieldnames = ['timestep', 'reward'] writer = csv.DictWriter(fd, fieldnames=fieldnames) if episode == 0: writer.writeheader() writer.writerow({'timestep': env.timestep, 'reward': reward}) # Close files in the logger logger.close_files() # Plot the learning curve logger.plot('PPO') # Save model save_dir = 'models/nolimit_holdem_ppo' if not os.path.exists(save_dir):
def nfsp(): import tensorflow as tf if tf.test.gpu_device_name(): print('GPU found') else: print("No GPU found") #os.environ['TF_CPP_MIN_LOG_LEVEL']='2' # Make environment env = rlcard.make('no-limit-holdem', config={ 'game_player_num': 2, 'seed': 477 }) eval_env = rlcard.make('no-limit-holdem', config={ 'seed': 12, 'game_player_num': 2 }) eval_env2 = rlcard.make('no-limit-holdem', config={ 'seed': 43, 'game_player_num': 2 }) #eval_env3 = rlcard.make('no-limit-holdem', config={'seed': 43, 'game_player_num': 2}) # Set the iterations numbers and how frequently we evaluate the performance # The intial memory size memory_init_size = 1000 # The paths for saving the logs and learning curves log_dir = './experiments/nolimit_holdem_nfsp_result/no_all_in' # Set a global seed set_global_seed(477) graph = tf.Graph() tf.ConfigProto() sess = tf.Session(graph=graph) evaluate_every = 2048 evaluate_num = 32 episode_num = 24576 # The intial memory size memory_init_size = 256 # Train the agent every X steps train_every = 256 agents = [] with graph.as_default(): """ def __init__(self, sess, scope, action_num=4, state_shape=None, hidden_layers_sizes=None, reservoir_buffer_capacity=int(1e6), anticipatory_param=0.1, batch_size=256, train_every=1, rl_learning_rate=0.1, sl_learning_rate=0.005, min_buffer_size_to_learn=1000, q_replay_memory_size=30000, q_replay_memory_init_size=1000, q_update_target_estimator_every=1000, q_discount_factor=0.99, q_epsilon_start=0.06, q_epsilon_end=0, q_epsilon_decay_steps=int(1e6), q_batch_size=256, q_train_every=1, q_mlp_layers=None, evaluate_with='average_policy'): """ # Model1v1V3cp10good agents.append( NFSPAgent(sess, scope='nfsp' + str(0), action_num=env.action_num, state_shape=env.state_shape, hidden_layers_sizes=[512, 512], anticipatory_param=0.1, rl_learning_rate=0.01, sl_learning_rate=0.005, q_epsilon_start=.7, min_buffer_size_to_learn=memory_init_size, q_replay_memory_size=80000, q_replay_memory_init_size=memory_init_size, train_every=train_every + 44, q_train_every=train_every, q_mlp_layers=[512, 512])) agents.append( NFSPAgent(sess, scope='nfsp' + str(1), action_num=env.action_num, state_shape=env.state_shape, hidden_layers_sizes=[512, 512], anticipatory_param=0.1, rl_learning_rate=0.01, sl_learning_rate=0.005, q_epsilon_start=.7, q_replay_memory_size=80000, min_buffer_size_to_learn=memory_init_size, q_replay_memory_init_size=memory_init_size, train_every=train_every + 44, q_train_every=train_every, q_mlp_layers=[512, 512])) # check_point_path = os.path.join('models\\nolimit_holdem_nfsp\\iivan') print( '-------------------------------------------------------------------------------------' ) # print(check_point_path) #todays project :) # https://stackoverflow.com/questions/33758669/running-multiple-tensorflow-sessions-concurrently with sess.as_default(): with graph.as_default(): # saver = tf.train.Saver() # saver.restore(sess, tf.train.latest_checkpoint(check_point_path)) global_step = tf.Variable(0, name='global_step', trainable=False) random_agent = RandomAgent(action_num=eval_env2.action_num) env.set_agents(agents) eval_env.set_agents([agents[0], random_agent]) eval_env2.set_agents([random_agent, agents[1]]) # eval_env3.set_agents([agents[1], random_agent]) # Initialize global variables sess.run(tf.global_variables_initializer()) # Init a Logger to plot the learning curve logger = Logger(log_dir) for episode in range(episode_num): print(episode, end='\r') #print('oh') # First sample a policy for the episode for agent in agents: agent.sample_episode_policy() # Generate data from the environment trajectories, _ = env.run(is_training=True) # Feed transitions into agent memory, and train the agent for i in range(env.player_num): for ts in trajectories[i]: agents[i].feed(ts) # Evaluate the performance. Play with random agents. if episode % evaluate_every == 0: logger.log( '\n\n\n---------------------------------------------------------------\nTournament ' + str(episode / evaluate_every)) # tournament(eval_env2, 6) # exploitability.exploitability(eval_env, agents[0], 500) res = tournament(env, evaluate_num) logger.log_performance(env.timestep, res[0]) res2 = tournament(eval_env, evaluate_num // 3) logger.log_performance(env.timestep, res2[0]) res3 = tournament(eval_env2, evaluate_num // 3) logger.log_performance(env.timestep, res3[0]) logger.log('' + str(episode_num) + " - " + str(episode) + '\n') logger.log( '\n\n----------------------------------------------------------------' ) if episode % (evaluate_every) == 0 and not episode == 0: save_dir = 'models/nolimit_holdem_nfsp/no_all_in/cp/' + str( episode // evaluate_every) if not os.path.exists(save_dir): os.makedirs(save_dir) saver = tf.train.Saver() saver.save(sess, os.path.join(save_dir, 'model')) logger.log( '\n\n\n---------------------------------------------------------------\nTournament ' + str(episode / evaluate_every)) res = tournament(eval_env, evaluate_num) logger.log_performance(env.timestep, res[0]) logger.log('' + str(episode_num) + " - " + str(episode)) # Close files in the logger logger.close_files() # Plot the learning curve logger.plot('NFSP') # Save model save_dir = 'models/nolimit_holdem_nfsp/no_all_in' if not os.path.exists(save_dir): os.makedirs(save_dir) saver = tf.train.Saver() saver.save(sess, os.path.join(save_dir, 'model'))
# Init a Logger to plot the learning curve logger_mcts = Logger(log_dir_mcts) logger_nfsp = Logger(log_dir_nfsp) for episode in range(episode_num): # Generate data from the environment trajectories, _ = env.run(is_training=True) # Feed transitions into agent memory, and train the agent for ts in trajectories[0]: nfsp_agent.feed(ts) # Evaluate the performance. Play with random agents. if episode % evaluate_every == 0: logger_mcts.log_performance(env.timestep, tournament(eval_env, evaluate_num)[0]) logger_nfsp.log_performance(env.timestep, tournament(eval_env, evaluate_num)[1]) # Close files in the logger logger_mcts.close_files() logger_nfsp.close_files() # Plot the learning curve logger_mcts.plot('MCTS') logger_nfsp.plot('NFSP') pd.DataFrame.to_csv(mcts_agent.action_df, os.path.join(log_dir_mcts, 'action.csv')) # Save model
for episode in range(episode_num): ## dont need these for random agent # Generate data from the environment #trajectories, _ = env.run(is_training=True) # Feed transitions into agent memory, and train the agent #for ts in trajectories[0]: # agent.feed(ts) # Evaluate the performance. Play with random agents. if episode % evaluate_every == 0: payoffs, peasant_wins, landlord_wins = tournament( eval_env, evaluate_num) logger.log_performance(episode, payoffs[0]) #print("DQN: ", peasant_wins, " and ", landlord_wins) logger.log_peasants(episode, peasant_wins / evaluate_num) logger.log_landlord(episode, landlord_wins / evaluate_num) # Close files in the logger logger.close_files() # Plot the learning curve logger.plot('Random', 'peasant_wins') logger.plot('Random', 'reward') logger.plot('Random', 'landlord_wins') # Save model nr = 0
eval_env = rlcard.make('doudizhu', config={ 'seed': 0, 'allow_step_back': True }) eval_env.set_agents([ agent, SRandomAgent(eval_env.action_num, seed=0), SRandomAgent(eval_env.action_num, seed=0) ]) time_start = time.time() payoffs1 = general_tournament(eval_env, evaluate_num, True) logger.log("episode:{} time:{} landlord winrate:{}".format( episode, time.time() - time_start, payoffs1[0])) L_WR_logger.log_performance(episode, payoffs1[0]) eval_env = rlcard.make('doudizhu', config={ 'seed': 0, 'allow_step_back': True }) eval_env.set_agents([ SRandomAgent(eval_env.action_num, seed=0), SRandomAgent(eval_env.action_num, seed=0), agent ]) time_start = time.time() payoffs2 = general_tournament(eval_env, evaluate_num, True) logger.log("episode:{} time:{} peasant winrate:{}".format( episode, time.time() - time_start, payoffs2[1]))
# Set a global seed set_global_seed(0) # Set up the agents agent = MPMCTSAgent(eval_env, emu_num=100) rdm_agent = RandomAgent(action_num=eval_env.action_num) eval_env.set_agents([agent, rdm_agent, rdm_agent]) eval_env.run(is_training=False) print(eval_env.game.round.trace) # Init a Logger to plot the learning curve logger = Logger(log_dir) for episode in range(episode_num): # Evaluate the performance. Play with random agents. if episode % evaluate_every == 0: logger.log_performance(eval_env.timestep, general_tournament(eval_env, evaluate_num)[0]) # Close files in the logger logger.close_files() # Plot the learning curve logger.plot('MCTS') # Save model save_dir = 'models/blackjack_mcts' if not os.path.exists(save_dir): os.makedirs(save_dir) saver = tf.train.Saver()
def main(): # Make environment env = rlcard.make('leduc-holdem', config={'seed': 0, 'env_num': 4}) eval_env = rlcard.make('leduc-holdem', config={'seed': 0, 'env_num': 4}) # Set the iterations numbers and how frequently we evaluate the performance evaluate_every = 100 evaluate_num = 10000 episode_num = 800000 # The intial memory size memory_init_size = 1000 # Train the agent every X steps train_every = 1 _reward_max = -0.5 # The paths for saving the logs and learning curves log_dir = './experiments/leduc_holdem_dqn_result/' # Set a global seed set_global_seed(0) with tf.Session() as sess: # Initialize a global step global_step = tf.Variable(0, name='global_step', trainable=False) # Set up the agents agent = DQNAgent(sess, scope='dqn', action_num=env.action_num, replay_memory_init_size=memory_init_size, train_every=train_every, state_shape=env.state_shape, mlp_layers=[128, 128]) # random_agent = RandomAgent(action_num=eval_env.action_num) cfr_agent = models.load('leduc-holdem-cfr').agents[0] env.set_agents([agent, agent]) eval_env.set_agents([agent, cfr_agent]) # Initialize global variables sess.run(tf.global_variables_initializer()) # Init a Logger to plot the learning curve logger = Logger(log_dir) saver = tf.train.Saver() save_dir = 'models/leduc_holdem_dqn' saver.restore(sess, os.path.join(save_dir, 'model')) for episode in range(episode_num): # Generate data from the environment trajectories, _ = env.run(is_training=True) # Feed transitions into agent memory, and train the agent for ts in trajectories[0]: agent.feed(ts) # Evaluate the performance. Play with random agents. if episode % evaluate_every == 0: _reward = tournament(eval_env, evaluate_num)[0] logger.log_performance(episode, _reward) if _reward > _reward_max: # Save model if not os.path.exists(save_dir): os.makedirs(save_dir) saver.save(sess, os.path.join(save_dir, 'model')) _reward_max = _reward # Close files in the logger logger.close_files() # Plot the learning curve logger.plot('DQN')
def main(): # Make environment env = rlcard.make('no-limit-holdem', config={ 'seed': 0, 'env_num': 16, 'game_player_num': 4 }) eval_env = rlcard.make('no-limit-holdem', config={ 'seed': 0, 'env_num': 16 }) # Set the iterations numbers and how frequently we evaluate the performance evaluate_every = 100 evaluate_num = 1000 episode_num = 200000 # The intial memory size memory_init_size = 1000 # Train the agent every X steps train_every = 1 _reward_max = -0.8 # The paths for saving the logs and learning curves log_dir = './experiments/nolimit_holdem_dqn_result/' # Set a global seed set_global_seed(0) with tf.Session() as sess: # Initialize a global step global_step = tf.Variable(0, name='global_step', trainable=False) # Set up the agents agent = DQNAgent(sess, scope='dqn', action_num=env.action_num, replay_memory_init_size=memory_init_size, train_every=train_every, state_shape=env.state_shape, mlp_layers=[512, 512]) agent2 = NFSPAgent(sess, scope='nfsp', action_num=env.action_num, state_shape=env.state_shape, hidden_layers_sizes=[512, 512], anticipatory_param=0.1, min_buffer_size_to_learn=memory_init_size, q_replay_memory_init_size=memory_init_size, train_every=64, q_train_every=64, q_mlp_layers=[512, 512]) # Initialize global variables sess.run(tf.global_variables_initializer()) save_dir = 'models/nolimit_holdem_dqn' saver = tf.train.Saver() #saver.restore(sess, os.path.join(save_dir, 'model')) random_agent = RandomAgent(action_num=eval_env.action_num) env.set_agents([agent, agent, agent2, random_agent]) eval_env.set_agents([agent, agent2]) # Init a Logger to plot the learning curve logger = Logger(log_dir) for episode in range(episode_num): agent2.sample_episode_policy() # Generate data from the environment trajectories, _ = env.run(is_training=True) # Feed transitions into agent memory, and train the agent for ts in trajectories[0]: agent.feed(ts) for ts in trajectories[2]: agent2.feed(ts) # Evaluate the performance. Play with random agents. if episode % evaluate_every == 0: _reward = tournament(eval_env, evaluate_num)[0] logger.log_performance(episode, _reward) if _reward > _reward_max: if not os.path.exists(save_dir): os.makedirs(save_dir) saver.save(sess, os.path.join(save_dir, 'model')) _reward_max = _reward # Close files in the logger logger.close_files() if not os.path.exists(save_dir): os.makedirs(save_dir) saver.save(sess, os.path.join(save_dir, 'model_final'))
logger = Logger(log_dir) # Create dir for results save_dir = 'models/thousand_schnapsen_deep_cfr3' if not os.path.exists(save_dir): os.makedirs(save_dir) saver = tf.train.Saver() best_win_rate = 0 for episode in range(episode_num): agents[0].train() # Evaluate the performance. Play with random agents. if episode % evaluate_every == 0: payoffs, wins = tournament(eval_env, evaluate_num) logger.log_performance(env.timestep, payoffs[0]) logger.csv_file.flush() win_rate = (wins[0] * 100) / evaluate_num print(f'Win rate: {win_rate}') # Save model if episode % save_every == 0 and win_rate > best_win_rate: best_win_rate = win_rate saver.save(sess, os.path.join(save_dir, 'model')) # Close files in the logger logger.close_files() # Plot the learning curve logger.plot('DeepCFR')
# Set a global seed set_global_seed(0) # Set up agents agent1 = limitholdem_rule_models.LimitholdemRuleAgentV1() agent2 = MCTS_Agent(action_num=env.action_num, duration=duration, exploration=explore, model_action=model_action, model_hand_rank=model_hand_rank) env.set_agents([agent2, agent1]) eval_env.set_agents([agent2, agent1]) for i in range(num_tournaments): logger.log_performance(i * 10, tournament(eval_env, evaluate_num)[0]) # for episode in range(episode_num): # # # Generate data from the environment # trajectories, _ = env.run(is_training=True) # # # print(trajectories) # # # Evaluate the performance. Play with random agents. # if episode % evaluate_every == 0: # logger.log_performance(env.timestep, tournament(eval_env, evaluate_num)[0]) # Close files in the logger logger.close_files()
def train_uno(): # Make environment env = rlcard.make("uno", config={"seed": 0}) eval_env = rlcard.make("uno", config={"seed": 0}) # Set the iterations numbers and how frequently we evaluate the performance evaluate_every = 100 evaluate_num = 1000 episode_num = 3000 # The intial memory size memory_init_size = 1000 # Train the agent every X steps train_every = 100 # The paths for saving the logs and learning curves log_dir = "./experiments/uno_results_dqn/" # Set a global seed set_global_seed(0) params = { "scope": "DQN-Agent", "num_actions": env.action_num, "replay_memory_size": memory_init_size, "num_states": env.state_shape, "discount_factor": 0.99, "epsilon_start": 1.0, "epsilon_end": 0.1, "epsilon_decay_steps": 20000, "batch_size": 32, "train_every": 1, "mlp_layers": [512, 512], "lr": 0.0005, } agent_conf = DQN_conf(**params) agent = DQN_agent(agent_conf) random_agent = RandomAgent(action_num=eval_env.action_num) env.set_agents([agent, random_agent]) eval_env.set_agents([agent, random_agent]) logger = Logger(log_dir) for episode in range(episode_num): # Generate data from the environment trajectories, _ = env.run(is_training=True) # Feed transitions into agent memory, and train the agent for ts in trajectories[0]: agent.feed(ts) # Evaluate the performance. Play with random agents. if episode % evaluate_every == 0: logger.log_performance(env.timestep, tournament(eval_env, evaluate_num)[0]) # Close files in the logger logger.close_files() # Plot the learning curve logger.plot("DQN UNO") # Save model save_dir = "models/uno_dqn_pytorch" if not os.path.exists(save_dir): os.makedirs(save_dir) state_dict = agent.get_state_dict() print(state_dict.keys()) torch.save(state_dict, os.path.join(save_dir, "model.pth"))
payoff_avg.append(payoffs[0]) # Feed transitions into agent memory, and train the agent for agent_id in [0, 1, 2, 3]: for ts in trajectories[agent_id]: rl_loss, sl_loss = env.agents[agent_id].feed(ts) if rl_loss != None: # and agent_id == 0: rl_loss_avg.append(rl_loss) if sl_loss != None: # and agent_id == 0: sl_loss_avg.append(sl_loss) t.set_description("rl: {}, sl: {}, payoff: {}, e: {}, rsv: {}".format( round(rl_loss_avg.get(), 3), round(sl_loss_avg.get(), 3), round(payoff_avg.get(), 3), round(env.agents[0].get_rl_epsilon(), 3), env.agents[0].get_reservoir_buffer_size()), refresh=True) # Evaluate the performance. Play with random agents. if episode % evaluate_every == evaluate_every - 1: logger.log_performance( env.timestep, tournament_tractor(eval_env, evaluate_num)[0]) saver.save(sess, os.path.join(save_dir, 'model')) # Close files in the logger logger.close_files() # Plot the learning curve logger.plot('NFSP')
def nfsp(): import tensorflow as tf if tf.test.gpu_device_name(): print('GPU found') else: print("No GPU found") #os.environ['TF_CPP_MIN_LOG_LEVEL']='2' # Make environment env = rlcard.make('no-limit-holdem', config={ 'record_action': False, 'game_player_num': 2 }) eval_env = rlcard.make('no-limit-holdem', config={ 'seed': 12, 'game_player_num': 2 }) eval_env2 = rlcard.make('no-limit-holdem', config={ 'seed': 43, 'game_player_num': 2 }) # Set the iterations numbers and how frequently we evaluate the performance # The intial memory size memory_init_size = 1000 # The paths for saving the logs and learning curves log_dir = './experiments/nolimit_holdem_nfsp_result/1v1MCNFSPv3' # Set a global seed set_global_seed(0) graph = tf.Graph() sess = tf.Session(graph=graph) evaluate_every = 1000 evaluate_num = 250 episode_num = 5000 # The intial memory size memory_init_size = 1500 # Train the agent every X steps train_every = 256 agents = [] with graph.as_default(): # Model1v1V3cp10good agents.append( NFSPAgent(sess, scope='nfsp' + str(0), action_num=env.action_num, state_shape=env.state_shape, hidden_layers_sizes=[512, 512], anticipatory_param=0.1, rl_learning_rate=.1, min_buffer_size_to_learn=memory_init_size, q_replay_memory_init_size=memory_init_size, train_every=train_every, q_train_every=train_every, q_mlp_layers=[512, 512])) agents.append( NFSPAgent(sess, scope='nfsp' + str(1), action_num=env.action_num, state_shape=env.state_shape, hidden_layers_sizes=[512, 512], anticipatory_param=0.075, rl_learning_rate=0.075, min_buffer_size_to_learn=memory_init_size, q_replay_memory_init_size=memory_init_size, train_every=train_every // 2, q_train_every=train_every // 2, q_mlp_layers=[512, 512])) # check_point_path = os.path.join('models\\nolimit_holdem_nfsp\\1v1MCNFSPv3\\cp\\10') print( '-------------------------------------------------------------------------------------' ) # print(check_point_path) with sess.as_default(): with graph.as_default(): saver = tf.train.Saver() # saver.restore(sess, tf.train.latest_checkpoint(check_point_path)) global_step = tf.Variable(0, name='global_step', trainable=False) random_agent = RandomAgent(action_num=eval_env2.action_num) #easy_agent = nfsp_agents[0] print(agents) # print(nfsp_agents) env.set_agents(agents) eval_env.set_agents(agents) eval_env2.set_agents([agents[0], random_agent]) # Initialize global variables sess.run(tf.global_variables_initializer()) # Init a Logger to plot the learning curve logger = Logger(log_dir) for episode in range(episode_num): # First sample a policy for the episode for agent in agents: agent.sample_episode_policy() table = [] # Generate data from the environment trajectories, _ = env.run(is_training=True) # Feed transitions into agent memory, and train the agent for i in range(env.player_num): for ts in trajectories[i]: agents[i].feed(ts, table) # Evaluate the performance. Play with random agents. if episode % evaluate_every == 0: logger.log( '\n\n\n---------------------------------------------------------------\nTournament ' + str(episode / evaluate_every)) res = tournament(eval_env, evaluate_num) res2 = tournament(eval_env2, evaluate_num // 4) logger.log_performance(env.timestep, res[0]) logger.log_performance(env.timestep, res2[0]) logger.log('' + str(episode_num) + " - " + str(episode) + '\n') logger.log( '\n\n----------------------------------------------------------------' ) if episode % (evaluate_every) == 0 and not episode == 0: save_dir = 'models/nolimit_holdem_nfsp/1v1MCNFSPv3/cp/10/good' + str( episode // evaluate_every) if not os.path.exists(save_dir): os.makedirs(save_dir) saver = tf.train.Saver() saver.save(sess, os.path.join(save_dir, 'model')) logger.log( '\n\n\n---------------------------------------------------------------\nTournament ' + str(episode / evaluate_every)) res = tournament(eval_env, evaluate_num) logger.log_performance(env.timestep, res[0]) logger.log('' + str(episode_num) + " - " + str(episode)) # Close files in the logger logger.close_files() # Plot the learning curve logger.plot('NFSP') # Save model save_dir = 'models/nolimit_holdem_nfsp/1v1MCNFSPv3/cp/10/good' if not os.path.exists(save_dir): os.makedirs(save_dir) saver = tf.train.Saver() saver.save(sess, os.path.join(save_dir, 'model'))
def main(): parser = createParser() namespace = parser.parse_args(sys.argv[1:]) #random seed random_seed = namespace.random_seed #names env_name = namespace.env_name env_num = 1 test_name = namespace.test_name dir_name = str(env_name) + '_a2c_' + str(test_name) + str(random_seed) # Set the iterations numbers and how frequently we evaluate/save plot evaluate_every = namespace.evaluate_every evaluate_num = namespace.evaluate_num episode_num = namespace.episode_num # Train the agent every X steps train_every = namespace.train_every save_every = namespace.save_every # Make environment env_rand = rlcard.make(env_name, config={'seed': random_seed}) env_ddqn = rlcard.make(env_name, config={'seed': random_seed}) env_qpg = rlcard.make(env_name, config={'seed': random_seed}) env_lstm = rlcard.make(env_name, config={'seed': random_seed}) env_lstmqpg = rlcard.make(env_name, config={'seed': random_seed}) eval_env = rlcard.make(env_name, config={'seed': random_seed}) # The paths for saving the logs and learning curves log_dir = './experiments/rl/' + dir_name + '_result' # Save model save_dir = 'models/rl/' + dir_name + '_result' # Set a global seed set_global_seed(random_seed) # Initialize a global step global_step = tf.Variable(0, name='global_step', trainable=False) # Set up the agents agent_rand = RandomAgent(action_num=eval_env.action_num) agent_ddqn = DDQNAgent( action_num=eval_env.action_num, state_shape=eval_env.state_shape, epsilon_decay_coef=math.pow(0.05 / 1, 1.0 / (episode_num // train_every)), ) agent_lstm = A2CLSTMAgent( action_num=eval_env.action_num, state_shape=eval_env.state_shape, trainble=False, discount_factor=0.95, critic_lstm_layers=[1, 512], critic_mlp_layers=[3, 512], critic_activation_func='tanh', critic_kernel_initializer='glorot_uniform', critic_learning_rate=0.001, critic_bacth_size=128, actor_lstm_layers=[1, 512], actor_mlp_layers=[3, 512], actor_activation_func='tanh', actor_kernel_initializer='glorot_uniform', actor_learning_rate=0.0001, actor_bacth_size=512, entropy_coef=0.5, entropy_decoy=math.pow(0.1 / 0.5, 1.0 / (episode_num // train_every)), max_grad_norm=1, ) agent_qpg = A2CQPGAgent( action_num=eval_env.action_num, state_shape=eval_env.state_shape, trainble=False, discount_factor=0.95, critic_mlp_layers=[4, 512], critic_activation_func='tanh', critic_kernel_initializer='glorot_uniform', critic_learning_rate=0.001, critic_bacth_size=128, actor_mlp_layers=[4, 512], actor_activation_func='tanh', actor_kernel_initializer='glorot_uniform', actor_learning_rate=0.0001, actor_bacth_size=512, entropy_coef=1, entropy_decoy=math.pow(0.05 / 1, 1.0 / (episode_num // train_every)), max_grad_norm=1, ) agent_lstmqpg = A2CLSTMQPGAgent( action_num=eval_env.action_num, state_shape=eval_env.state_shape, trainable=False, discount_factor=0.95, critic_lstm_layers=[1, 512], critic_mlp_layers=[3, 512], critic_activation_func='tanh', critic_kernel_initializer='glorot_uniform', critic_learning_rate=0.001, critic_bacth_size=128, actor_lstm_layers=[1, 512], actor_mlp_layers=[3, 512], actor_activation_func='tanh', actor_kernel_initializer='glorot_uniform', actor_learning_rate=0.0001, actor_bacth_size=512, entropy_coef=0.5, entropy_decoy=math.pow(0.1 / 0.5, 1.0 / (episode_num // train_every)), max_grad_norm=1, ) agent_test = A2CLSTMAgent( action_num=eval_env.action_num, state_shape=eval_env.state_shape, discount_factor=0.95, critic_lstm_layers=[1, 512], critic_mlp_layers=[3, 512], critic_activation_func='tanh', critic_kernel_initializer='glorot_uniform', critic_learning_rate=0.001, critic_bacth_size=128, actor_lstm_layers=[1, 512], actor_mlp_layers=[3, 512], actor_activation_func='tanh', actor_kernel_initializer='glorot_uniform', actor_learning_rate=0.0001, actor_bacth_size=512, entropy_coef=0.5, entropy_decoy=math.pow(0.1 / 0.5, 1.0 / (episode_num // train_every)), max_grad_norm=1, ) if namespace.load_model is not None: agent_test.load_model(namespace.load_model) agent_ddqn.load_model('models/rl/no_limit_holdem_ddqn_result/test0') agent_lstm.load_model( 'models/rl/no_limit_holdem_a2c_v2_lstm_result/test1000') agent_qpg.load_model( 'models/rl/no_limit_holdem_a2c_v2_qpg_result/test1000') agent_lstmqpg.load_model( 'models/rl/no_limit_holdem_a2c_v2_lstm_qpg_result/test_r_900000') env_rand.set_agents([agent_test, agent_rand]) env_ddqn.set_agents([agent_test, agent_ddqn]) env_qpg.set_agents([agent_test, agent_qpg]) env_lstm.set_agents([agent_test, agent_lstm]) env_lstmqpg.set_agents([agent_test, agent_lstmqpg]) eval_env.set_agents([agent_test, agent_rand]) # Init a Logger to plot the learning curve logger = Logger(log_dir + '/' + test_name) envs = [env_rand, env_ddqn, env_qpg, env_lstm, env_lstmqpg] env_num = len(envs) for episode in range(episode_num // env_num): # Generate data from the for env in envs: trajectories, _ = env.run(is_training=True) #agent_test.reset_lstm_memory() # Feed transitions into agent memory, and train the agent for ts in trajectories[0]: agent_test.feed(ts) if episode % (train_every // env_num) == 0: agent_test.train() if episode % (save_every // env_num) == 0: # Save model if not os.path.exists(save_dir + '/' + test_name + str(episode * env_num)): os.makedirs(save_dir + '/' + test_name + str(episode * env_num)) agent_test.save_model(save_dir + '/' + test_name + str(episode * env_num)) # Evaluate the performance. Play with random agents. if episode % (evaluate_every // env_num) == 0: print('episode: ', episode * env_num) logger.log_performance(episode * env_num, tournament(eval_env, evaluate_num)[0]) # Close files in the logger logger.close_files() # Plot the learning curve logger.plot(dir_name) # Save model if not os.path.exists(save_dir + '/' + test_name + str(episode_num)): os.makedirs(save_dir + '/' + test_name + str(episode_num)) agent_test.save_model(save_dir + '/' + test_name + str(episode_num))
# Initialize global variables sess.run(tf.global_variables_initializer()) # restore checkpoint saver = tf.train.Saver() save_dir = 'models/nolimit_holdem_deepcfr' # Init a Logger to plot the learning curve logger = Logger(log_dir) for episode in range(episode_num): for agent in agents: agent.train() # Evaluate the performance. Play with random agents. if episode % evaluate_every == 0: _reward = tournament(eval_env, evaluate_num)[0] logger.log_performance(episode, _reward) # Save model if _reward > _reward_max: if not os.path.exists(save_dir): os.makedirs(save_dir) saver.save(sess, os.path.join(save_dir, 'model')) _reward_max = _reward # Close files in the logger logger.close_files() # Plot the learning curve logger.plot('DeepCFR')
for episode in range(episode_num): # Generate data from the environment trajectories, _ = env.run(is_training=True) # Feed transitions into agent memory, and train the agent for ts in trajectories[0]: agent.feed(ts) # Evaluate the performance. Play with random agents. if episode % evaluate_every == 0: ##payoffs, peasant_wins, landlord_wins = tournament(eval_env, evaluate_num) ## new with loss: payoffs, peasant_wins, landlord_wins, agent_peasant_wins, agent_landlord_wins = tournament(eval_env, evaluate_num) logger.log_performance(episode, payoffs[role_counter]) #print("DQN: ", peasant_wins, " and ", landlord_wins) logger.log_peasants(episode, peasant_wins/evaluate_num) logger.log_landlord(episode, landlord_wins/evaluate_num) logger.log_loss(episode, agent.get_loss()) logger.log_agent_peasant(episode, agent_peasant_wins) logger.log_agent_landlord(episode, agent_landlord_wins) # Close files in the logger logger.close_files() # Plot the learning curve logger.plot('DQN', 'peasant_wins') logger.plot('DQN', 'reward') logger.plot('DQN', 'landlord_wins')
def main(): # Make environment env = rlcard.make('blackjack', config={'env_num': 4, 'seed': 0}) eval_env = rlcard.make('blackjack', config={'env_num': 4, 'seed': 0}) # Set the iterations numbers and how frequently we evaluate performance evaluate_every = 100 evaluate_num = 10000 iteration_num = 100000 # The intial memory size memory_init_size = 100 # Train the agent every X steps train_every = 1 # The paths for saving the logs and learning curves log_dir = './experiments/blackjack_dqn_result/' # Set a global seed set_global_seed(0) with tf.Session() as sess: # Initialize a global step global_step = tf.Variable(0, name='global_step', trainable=False) # Set up the agents agent = DQNAgent(sess, scope='dqn', action_num=env.action_num, replay_memory_init_size=memory_init_size, train_every=train_every, state_shape=env.state_shape, mlp_layers=[10, 10]) env.set_agents([agent]) eval_env.set_agents([agent]) # Initialize global variables sess.run(tf.global_variables_initializer()) # Initialize a Logger to plot the learning curve logger = Logger(log_dir) for iteration in range(iteration_num): # Generate data from the environment trajectories, _ = env.run(is_training=True) # Feed transitions into agent memory, and train the agent for ts in trajectories[0]: agent.feed(ts) # Evaluate the performance. Play with random agents. if iteration % evaluate_every == 0: logger.log_performance(env.timestep, tournament(eval_env, evaluate_num)[0]) # Close files in the logger logger.close_files() # Plot the learning curve logger.plot('DQN') # Save model save_dir = 'models/blackjack_dqn' if not os.path.exists(save_dir): os.makedirs(save_dir) saver = tf.train.Saver() saver.save(sess, os.path.join(save_dir, 'model'))