def __init__(self, in_policy_file, out_policy_file, domainString='CamRestaurants', is_training=False):
super(BDQNPolicy, self).__init__(domainString, is_training)
tf.reset_default_graph()
self.domainString = domainString
self.domainUtil = FlatOnt.FlatDomainOntology(self.domainString)
self.in_policy_file = in_policy_file
self.out_policy_file = out_policy_file
self.is_training = is_training
self.accum_belief = []
self.domainUtil = FlatOnt.FlatDomainOntology(self.domainString)
self.prev_state_check = None
# parameter settings
self.n_in = 260
if cfg.has_option('dqnpolicy', 'n_in'):
self.n_in = cfg.getint('dqnpolicy', 'n_in')
self.actor_lr = 0.0001
if cfg.has_option('dqnpolicy', 'actor_lr'):
self.actor_lr = cfg.getfloat('dqnpolicy', 'actor_lr')
self.critic_lr = 0.001
if cfg.has_option('dqnpolicy', 'critic_lr'):
self.critic_lr = cfg.getfloat('dqnpolicy', 'critic_lr')
self.tau = 0.001
if cfg.has_option('dqnpolicy', 'tau'):
self.tau = cfg.getfloat('dqnpolicy', 'tau')
self.randomseed = 1234
if cfg.has_option('GENERAL', 'seed'):
self.randomseed = cfg.getint('GENERAL', 'seed')
self.gamma = 1.0
if cfg.has_option('dqnpolicy', 'gamma'):
self.gamma = cfg.getfloat('dqnpolicy', 'gamma')
self.regularisation = 'l2'
if cfg.has_option('dqnpolicy', 'regularisation'):
self.regularisation = cfg.get('dqnpolicy', 'regulariser')
self.learning_rate = 0.001 # ct506 #0.001
if cfg.has_option('dqnpolicy', 'learning_rate'):
self.learning_rate = cfg.getfloat('dqnpolicy', 'learning_rate')
self.exploration_type = 'e-greedy' # Boltzman
if cfg.has_option('dqnpolicy', 'exploration_type'):
self.exploration_type = cfg.get('dqnpolicy', 'exploration_type')
self.episodeNum = 1000
if cfg.has_option('dqnpolicy', 'episodeNum'):
self.episodeNum = cfg.getfloat('dqnpolicy', 'episodeNum')
self.maxiter = 5000
if cfg.has_option('dqnpolicy', 'maxiter'):
self.maxiter = cfg.getfloat('dqnpolicy', 'maxiter')
self.epsilon = 1
if cfg.has_option('dqnpolicy', 'epsilon'):
self.epsilon = cfg.getfloat('dqnpolicy', 'epsilon')
self.epsilon_start = 1
if cfg.has_option('dqnpolicy', 'epsilon_start'):
self.epsilon_start = cfg.getfloat('dqnpolicy', 'epsilon_start')
self.epsilon_end = 1
if cfg.has_option('dqnpolicy', 'epsilon_end'):
self.epsilon_end = cfg.getfloat('dqnpolicy', 'epsilon_end')
self.priorProbStart = 1.0
if cfg.has_option('dqnpolicy', 'prior_sample_prob_start'):
self.priorProbStart = cfg.getfloat('dqnpolicy', 'prior_sample_prob_start')
self.priorProbEnd = 0.1
if cfg.has_option('dqnpolicy', 'prior_sample_prob_end'):
self.priorProbEnd = cfg.getfloat('dqnpolicy', 'prior_sample_prob_end')
self.policyfeatures = []
if cfg.has_option('dqnpolicy', 'features'):
logger.info('Features: ' + str(cfg.get('dqnpolicy', 'features')))
self.policyfeatures = json.loads(cfg.get('dqnpolicy', 'features'))
self.max_k = 5
if cfg.has_option('dqnpolicy', 'max_k'):
self.max_k = cfg.getint('dqnpolicy', 'max_k')
self.learning_algorithm = 'drl'
if cfg.has_option('dqnpolicy', 'learning_algorithm'):
self.learning_algorithm = cfg.get('dqnpolicy', 'learning_algorithm')
logger.info('Learning algorithm: ' + self.learning_algorithm)
self.minibatch_size = 32
if cfg.has_option('dqnpolicy', 'minibatch_size'):
self.minibatch_size = cfg.getint('dqnpolicy', 'minibatch_size')
self.capacity = 1000 # max(self.minibatch_size, 2000)
if cfg.has_option('dqnpolicy', 'capacity'):
self.capacity = max(cfg.getint('dqnpolicy', 'capacity'), 2000)
self.replay_type = 'vanilla'
if cfg.has_option('dqnpolicy', 'replay_type'):
self.replay_type = cfg.get('dqnpolicy', 'replay_type')
self.architecture = 'vanilla'
if cfg.has_option('dqnpolicy', 'architecture'):
self.architecture = cfg.get('dqnpolicy', 'architecture')
self.q_update = 'single'
if cfg.has_option('dqnpolicy', 'q_update'):
self.q_update = cfg.get('dqnpolicy', 'q_update')
self.h1_size = 130
if cfg.has_option('dqnpolicy', 'h1_size'):
self.h1_size = cfg.getint('dqnpolicy', 'h1_size')
self.h2_size = 130
if cfg.has_option('dqnpolicy', 'h2_size'):
self.h2_size = cfg.getint('dqnpolicy', 'h2_size')
self.save_step = 200
if cfg.has_option('policy', 'save_step'):
self.save_step = cfg.getint('policy', 'save_step')
# BDQN parameteres
self.n_samples = 1
if cfg.has_option('dqnpolicy', 'n_samples'):
self.n_samples = cfg.getint('dqnpolicy', 'n_samples')
sigma_prior = 1.5 # np.array(-3.0, dtype=np.float32)
if cfg.has_option('dqnpolicy', 'sigma_prior'):
sigma_prior = cfg.getfloat('dqnpolicy', 'sigma_prior')
self.sigma_prior = tf.exp(sigma_prior) # np.exp(np.array(sigma_prior, dtype=np.float32))
self.stddev_var_mu = 0.01
if cfg.has_option('dqnpolicy', 'stddev_var_mu'):
self.stddev_var_mu = cfg.getfloat('dqnpolicy', 'stddev_var_mu')
self.stddev_var_logsigma = 0.01
if cfg.has_option('dqnpolicy', 'stddev_var_logsigma'):
self.stddev_var_logsigma = cfg.getfloat('dqnpolicy', 'stddev_var_logsigma')
self.mean_log_sigma = 0.000001
if cfg.has_option('dqnpolicy', 'mean_log_sigma'):
self.mean_log_sigma = cfg.getfloat('dqnpolicy', 'mean_log_sigma')
self.n_batches = 1000.0
if cfg.has_option('dqnpolicy', 'n_batches'):
self.n_batches = cfg.getfloat('dqnpolicy', 'n_batches')
self.importance_sampling = False
if cfg.has_option('dqnpolicy', 'importance_sampling'):
self.importance_sampling = cfg.getboolean('dqnpolicy', 'importance_sampling')
self.alpha = 0.85
if cfg.has_option('dqnpolicy', 'alpha'):
self.alpha = cfg.getfloat('dqnpolicy', 'alpha')
self.alpha_divergence = False
if cfg.has_option('dqnpolicy', 'alpha_divergence'):
self.alpha_divergence = cfg.getboolean('dqnpolicy', 'alpha_divergence')
self.sigma_eps = 0.01
if cfg.has_option('dqnpolicy', 'sigma_eps'):
self.sigma_eps = cfg.getfloat('dqnpolicy', 'sigma_eps')
self.training_frequency = 2
if cfg.has_option('dqnpolicy', 'training_frequency'):
self.training_frequency = cfg.getint('dqnpolicy', 'training_frequency')
# domain specific parameter settings (overrides general policy parameter settings)
if cfg.has_option('dqnpolicy_' + domainString, 'n_in'):
self.n_in = cfg.getint('dqnpolicy_' + domainString, 'n_in')
if cfg.has_option('dqnpolicy_' + domainString, 'actor_lr'):
self.actor_lr = cfg.getfloat('dqnpolicy_' + domainString, 'actor_lr')
if cfg.has_option('dqnpolicy_' + domainString, 'critic_lr'):
self.critic_lr = cfg.getfloat('dqnpolicy_' + domainString, 'critic_lr')
if cfg.has_option('dqnpolicy_' + domainString, 'tau'):
self.tau = cfg.getfloat('dqnpolicy_' + domainString, 'tau')
if cfg.has_option('dqnpolicy_' + domainString, 'gamma'):
self.gamma = cfg.getfloat('dqnpolicy_' + domainString, 'gamma')
if cfg.has_option('dqnpolicy_' + domainString, 'regularisation'):
self.regularisation = cfg.get('dqnpolicy_' + domainString, 'regulariser')
if cfg.has_option('dqnpolicy_' + domainString, 'learning_rate'):
self.learning_rate = cfg.getfloat('dqnpolicy_' + domainString, 'learning_rate')
if cfg.has_option('dqnpolicy_' + domainString, 'exploration_type'):
self.exploration_type = cfg.get('dqnpolicy_' + domainString, 'exploration_type')
if cfg.has_option('dqnpolicy_' + domainString, 'episodeNum'):
self.episodeNum = cfg.getfloat('dqnpolicy_' + domainString, 'episodeNum')
if cfg.has_option('dqnpolicy_' + domainString, 'maxiter'):
self.maxiter = cfg.getfloat('dqnpolicy_' + domainString, 'maxiter')
if cfg.has_option('dqnpolicy_' + domainString, 'epsilon'):
self.epsilon = cfg.getfloat('dqnpolicy_' + domainString, 'epsilon')
if cfg.has_option('dqnpolicy_' + domainString, 'epsilon_start'):
self.epsilon_start = cfg.getfloat('dqnpolicy_' + domainString, 'epsilon_start')
if cfg.has_option('dqnpolicy_' + domainString, 'epsilon_end'):
self.epsilon_end = cfg.getfloat('dqnpolicy_' + domainString, 'epsilon_end')
if cfg.has_option('dqnpolicy_' + domainString, 'prior_sample_prob_start'):
self.priorProbStart = cfg.getfloat('dqnpolicy_' + domainString, 'prior_sample_prob_start')
if cfg.has_option('dqnpolicy_' + domainString, 'prior_sample_prob_end'):
self.priorProbEnd = cfg.getfloat('dqnpolicy_' + domainString, 'prior_sample_prob_end')
if cfg.has_option('dqnpolicy_' + domainString, 'features'):
logger.info('Features: ' + str(cfg.get('dqnpolicy_' + domainString, 'features')))
self.policyfeatures = json.loads(cfg.get('dqnpolicy_' + domainString, 'features'))
if cfg.has_option('dqnpolicy_' + domainString, 'max_k'):
self.max_k = cfg.getint('dqnpolicy_' + domainString, 'max_k')
if cfg.has_option('dqnpolicy_' + domainString, 'learning_algorithm'):
self.learning_algorithm = cfg.get('dqnpolicy_' + domainString, 'learning_algorithm')
logger.info('Learning algorithm: ' + self.learning_algorithm)
if cfg.has_option('dqnpolicy_' + domainString, 'minibatch_size'):
self.minibatch_size = cfg.getint('dqnpolicy_' + domainString, 'minibatch_size')
if cfg.has_option('dqnpolicy_' + domainString, 'capacity'):
self.capacity = cfg.getint('dqnpolicy_' + domainString, 'capacity')
if cfg.has_option('dqnpolicy_' + domainString, 'replay_type'):
self.replay_type = cfg.get('dqnpolicy_' + domainString, 'replay_type')
if cfg.has_option('dqnpolicy_' + domainString, 'architecture'):
self.architecture = cfg.get('dqnpolicy_' + domainString, 'architecture')
if cfg.has_option('dqnpolicy_' + domainString, 'q_update'):
self.q_update = cfg.get('dqnpolicy_' + domainString, 'q_update')
if cfg.has_option('dqnpolicy_' + domainString, 'h1_size'):
self.h1_size = cfg.getint('dqnpolicy_' + domainString, 'h1_size')
if cfg.has_option('dqnpolicy_' + domainString, 'h2_size'):
self.h2_size = cfg.getint('dqnpolicy_' + domainString, 'h2_size')
if cfg.has_option('policy_' + domainString, 'save_step'):
self.save_step = cfg.getint('policy_' + domainString, 'save_step')
# BDQN parameteres
if cfg.has_option('dqnpolicy_' + domainString, 'n_samples'):
self.n_samples = cfg.getint('dqnpolicy_' + domainString, 'n_samples')
if cfg.has_option('dqnpolicy_' + domainString, 'sigma_prior'):
sigma_prior = cfg.getfloat('dqnpolicy_' + domainString, 'sigma_prior')
self.sigma_prior = tf.exp(sigma_prior) # np.exp(np.array(sigma_prior, dtype=np.float32))
if cfg.has_option('dqnpolicy_' + domainString, 'stddev_var_mu'):
self.stddev_var_mu = cfg.getfloat('dqnpolicy_' + domainString, 'stddev_var_mu')
if cfg.has_option('dqnpolicy_' + domainString, 'stddev_var_logsigma'):
self.stddev_var_logsigma = cfg.getfloat('dqnpolicy_' + domainString, 'stddev_var_logsigma')
if cfg.has_option('dqnpolicy_' + domainString, 'mean_log_sigma'):
self.mean_log_sigma = cfg.getfloat('dqnpolicy_' + domainString, 'mean_log_sigma')
if cfg.has_option('dqnpolicy_' + domainString, 'n_batches'):
self.n_batches = cfg.getfloat('dqnpolicy_' + domainString, 'n_batches')
if cfg.has_option('dqnpolicy_' + domainString, 'importance_sampling'):
self.importance_sampling = cfg.getboolean('dqnpolicy_' + domainString, 'importance_sampling')
if cfg.has_option('dqnpolicy_' + domainString, 'alpha'):
self.alpha = cfg.getfloat('dqnpolicy_' + domainString, 'alpha')
if cfg.has_option('dqnpolicy_' + domainString, 'alpha_divergence'):
self.alpha_divergence = cfg.getboolean('dqnpolicy_' + domainString, 'alpha_divergence')
if cfg.has_option('dqnpolicy_' + domainString, 'sigma_eps'):
self.sigma_eps = cfg.getfloat('dqnpolicy_' + domainString, 'sigma_eps')
if cfg.has_option('dqnpolicy_' + domainString, 'training_frequency'):
self.training_frequency = cfg.getint('dqnpolicy_' + domainString, 'training_frequency')
print 'ct506', 'sigma_eps', self.sigma_eps, 'lr', self.learning_rate, 'm', self.n_batches
self.episode_ave_max_q = []
os.environ["CUDA_VISIBLE_DEVICES"] = ""
# init session
self.sess = tf.Session()
with tf.device("/cpu:0"):
np.random.seed(self.randomseed)
tf.set_random_seed(self.randomseed)
# initialise an replay buffer
if self.replay_type == 'vanilla':
self.episodes[self.domainString] = ReplayBuffer(self.capacity, self.minibatch_size, self.randomseed)
elif self.replay_type == 'prioritized':
self.episodes[self.domainString] = ReplayPrioritised(self.capacity, self.minibatch_size,
self.randomseed)
# replay_buffer = ReplayBuffer(self.capacity, self.randomseed)
# self.episodes = []
self.samplecount = 0
self.episodecount = 0
# construct the models
self.state_dim = self.n_in
self.summaryaction = SummaryAction.SummaryAction(domainString)
self.action_dim = len(self.summaryaction.action_names)
action_bound = len(self.summaryaction.action_names)
self.stats = [0 for _ in range(self.action_dim)]
self.stdVar = []
self.meanVar = []
self.stdMean = []
self.meanMean = []
self.td_error = []
self.td_errorVar = []
self.bbqn = bbqn.DeepQNetwork(self.sess, self.state_dim, self.action_dim, self.learning_rate, self.tau, action_bound, self.architecture, self.h1_size, self.h2_size, self.n_samples, self.minibatch_size, self.sigma_prior, self.n_batches, self.stddev_var_mu, self.stddev_var_logsigma, self.mean_log_sigma, self.importance_sampling, self.alpha_divergence, self.alpha, self.sigma_eps)
# when all models are defined, init all variables
init_op = tf.global_variables_initializer()
self.sess.run(init_op)
self.loadPolicy(self.in_policy_file)
print 'loaded replay size: ', self.episodes[self.domainString].size()
self.bbqn.update_target_network()
def __init__(self, in_policy_file, out_policy_file, domainString='CamRestaurants', is_training=False):
super(ENACPolicy, self).__init__(domainString, is_training)
tf.reset_default_graph()
self.in_policy_file = in_policy_file
self.out_policy_file = out_policy_file
self.is_training = is_training
self.accum_belief = []
self.prev_state_check = None
self.domainString = domainString
self.domainUtil = FlatOnt.FlatDomainOntology(self.domainString)
# parameter settings
if 0:#cfg.has_option('dqnpolicy', 'n_in'): #ic304: this was giving me a weird error, disabled it until i can check it deeper
self.n_in = cfg.getint('dqnpolicy', 'n_in')
else:
self.n_in = self.get_n_in(domainString)
self.actor_lr = 0.0001
if cfg.has_option('dqnpolicy', 'actor_lr'):
self.actor_lr = cfg.getfloat('dqnpolicy', 'actor_lr')
self.critic_lr = 0.001
if cfg.has_option('dqnpolicy', 'critic_lr'):
self.critic_lr = cfg.getfloat('dqnpolicy', 'critic_lr')
self.tau = 0.001
if cfg.has_option('dqnpolicy', 'tau'):
self.tau = cfg.getfloat('dqnpolicy', 'tau')
self.randomseed = 1234
if cfg.has_option('GENERAL', 'seed'):
self.randomseed = cfg.getint('GENERAL', 'seed')
self.gamma = 1.0
if cfg.has_option('dqnpolicy', 'gamma'):
self.gamma = cfg.getfloat('dqnpolicy', 'gamma')
self.regularisation = 'l2'
if cfg.has_option('dqnpolicy', 'regularisation'):
self.regularisation = cfg.get('dqnpolicy', 'regulariser')
self.learning_rate = 0.001
if cfg.has_option('dqnpolicy', 'learning_rate'):
self.learning_rate = cfg.getfloat('dqnpolicy', 'learning_rate')
self.exploration_type = 'e-greedy' # Boltzman
if cfg.has_option('dqnpolicy', 'exploration_type'):
self.exploration_type = cfg.get('dqnpolicy', 'exploration_type')
self.episodeNum = 1000
if cfg.has_option('dqnpolicy', 'episodeNum'):
self.episodeNum = cfg.getfloat('dqnpolicy', 'episodeNum')
self.maxiter = 5000
if cfg.has_option('dqnpolicy', 'maxiter'):
self.maxiter = cfg.getfloat('dqnpolicy', 'maxiter')
self.epsilon = 1
if cfg.has_option('dqnpolicy', 'epsilon'):
self.epsilon = cfg.getfloat('dqnpolicy', 'epsilon')
self.epsilon_start = 1
if cfg.has_option('dqnpolicy', 'epsilon_start'):
self.epsilon_start = cfg.getfloat('dqnpolicy', 'epsilon_start')
self.epsilon_end = 1
if cfg.has_option('dqnpolicy', 'epsilon_end'):
self.epsilon_end = cfg.getfloat('dqnpolicy', 'epsilon_end')
self.priorProbStart = 1.0
if cfg.has_option('dqnpolicy', 'prior_sample_prob_start'):
self.priorProbStart = cfg.getfloat('dqnpolicy', 'prior_sample_prob_start')
self.priorProbEnd = 0.1
if cfg.has_option('dqnpolicy', 'prior_sample_prob_end'):
self.priorProbEnd = cfg.getfloat('dqnpolicy', 'prior_sample_prob_end')
self.policyfeatures = []
if cfg.has_option('dqnpolicy', 'features'):
logger.info('Features: ' + str(cfg.get('dqnpolicy', 'features')))
self.policyfeatures = json.loads(cfg.get('dqnpolicy', 'features'))
self.max_k = 5
if cfg.has_option('dqnpolicy', 'max_k'):
self.max_k = cfg.getint('dqnpolicy', 'max_k')
self.learning_algorithm = 'drl'
if cfg.has_option('dqnpolicy', 'learning_algorithm'):
self.learning_algorithm = cfg.get('dqnpolicy', 'learning_algorithm')
logger.info('Learning algorithm: ' + self.learning_algorithm)
self.minibatch_size = 32
if cfg.has_option('dqnpolicy', 'minibatch_size'):
self.minibatch_size = cfg.getint('dqnpolicy', 'minibatch_size')
self.capacity = 1000 # max(self.minibatch_size, 2000)
if cfg.has_option('dqnpolicy', 'capacity'):
self.capacity = cfg.getint('dqnpolicy', 'capacity')
self.replay_type = 'vanilla'
if cfg.has_option('dqnpolicy', 'replay_type'):
self.replay_type = cfg.get('dqnpolicy', 'replay_type')
self.architecture = 'vanilla'
if cfg.has_option('dqnpolicy', 'architecture'):
self.architecture = cfg.get('dqnpolicy', 'architecture')
self.q_update = 'single'
if cfg.has_option('dqnpolicy', 'q_update'):
self.q_update = cfg.get('dqnpolicy', 'q_update')
self.h1_size = 130
if cfg.has_option('dqnpolicy', 'h1_size'):
self.h1_size = cfg.getint('dqnpolicy', 'h1_size')
self.h2_size = 50
if cfg.has_option('dqnpolicy', 'h2_size'):
self.h2_size = cfg.getint('dqnpolicy', 'h2_size')
self.save_step = 200
if cfg.has_option('policy', 'save_step'):
self.save_step = cfg.getint('policy', 'save_step')
self.importance_sampling = 'soft'
if cfg.has_option('dqnpolicy', 'importance_sampling'):
self.importance_sampling = cfg.get('dqnpolicy', 'importance_sampling')
self.training_frequency = 2
if cfg.has_option('dqnpolicy', 'training_frequency'):
self.training_frequency = cfg.getint('dqnpolicy', 'training_frequency')
# domain specific parameter settings (overrides general policy parameter settings)
if cfg.has_option('dqnpolicy_' + domainString, 'n_in'):
self.n_in = cfg.getint('dqnpolicy_' + domainString, 'n_in')
if cfg.has_option('dqnpolicy_' + domainString, 'actor_lr'):
self.actor_lr = cfg.getfloat('dqnpolicy_' + domainString, 'actor_lr')
if cfg.has_option('dqnpolicy_' + domainString, 'critic_lr'):
self.critic_lr = cfg.getfloat('dqnpolicy_' + domainString, 'critic_lr')
if cfg.has_option('dqnpolicy_' + domainString, 'tau'):
self.tau = cfg.getfloat('dqnpolicy_' + domainString, 'tau')
if cfg.has_option('dqnpolicy_' + domainString, 'gamma'):
self.gamma = cfg.getfloat('dqnpolicy_' + domainString, 'gamma')
if cfg.has_option('dqnpolicy_' + domainString, 'regularisation'):
self.regularisation = cfg.get('dqnpolicy_' + domainString, 'regulariser')
if cfg.has_option('dqnpolicy_' + domainString, 'learning_rate'):
self.learning_rate = cfg.getfloat('dqnpolicy_' + domainString, 'learning_rate')
if cfg.has_option('dqnpolicy_' + domainString, 'exploration_type'):
self.exploration_type = cfg.get('dqnpolicy_' + domainString, 'exploration_type')
if cfg.has_option('dqnpolicy_' + domainString, 'episodeNum'):
self.episodeNum = cfg.getfloat('dqnpolicy_' + domainString, 'episodeNum')
if cfg.has_option('dqnpolicy_' + domainString, 'maxiter'):
self.maxiter = cfg.getfloat('dqnpolicy_' + domainString, 'maxiter')
if cfg.has_option('dqnpolicy_' + domainString, 'epsilon'):
self.epsilon = cfg.getfloat('dqnpolicy_' + domainString, 'epsilon')
if cfg.has_option('dqnpolicy_' + domainString, 'epsilon_start'):
self.epsilon_start = cfg.getfloat('dqnpolicy_' + domainString, 'epsilon_start')
if cfg.has_option('dqnpolicy_' + domainString, 'epsilon_end'):
self.epsilon_end = cfg.getfloat('dqnpolicy_' + domainString, 'epsilon_end')
if cfg.has_option('dqnpolicy_' + domainString, 'prior_sample_prob_start'):
self.priorProbStart = cfg.getfloat('dqnpolicy_' + domainString, 'prior_sample_prob_start')
if cfg.has_option('dqnpolicy_' + domainString, 'prior_sample_prob_end'):
self.priorProbEnd = cfg.getfloat('dqnpolicy_' + domainString, 'prior_sample_prob_end')
if cfg.has_option('dqnpolicy_' + domainString, 'features'):
logger.info('Features: ' + str(cfg.get('dqnpolicy_' + domainString, 'features')))
self.policyfeatures = json.loads(cfg.get('dqnpolicy_' + domainString, 'features'))
if cfg.has_option('dqnpolicy_' + domainString, 'max_k'):
self.max_k = cfg.getint('dqnpolicy_' + domainString, 'max_k')
self.learning_algorithm = 'drl'
if cfg.has_option('dqnpolicy_' + domainString, 'learning_algorithm'):
self.learning_algorithm = cfg.get('dqnpolicy_' + domainString, 'learning_algorithm')
logger.info('Learning algorithm: ' + self.learning_algorithm)
if cfg.has_option('dqnpolicy_' + domainString, 'minibatch_size'):
self.minibatch_size = cfg.getint('dqnpolicy_' + domainString, 'minibatch_size')
if cfg.has_option('dqnpolicy_' + domainString, 'capacity'):
self.capacity = cfg.getint('dqnpolicy_' + domainString, 'capacity')
if cfg.has_option('dqnpolicy_' + domainString, 'replay_type'):
self.replay_type = cfg.get('dqnpolicy_' + domainString, 'replay_type')
if cfg.has_option('dqnpolicy_' + domainString, 'architecture'):
self.architecture = cfg.get('dqnpolicy_' + domainString, 'architecture')
if cfg.has_option('dqnpolicy_' + domainString, 'q_update'):
self.q_update = cfg.get('dqnpolicy_' + domainString, 'q_update')
if cfg.has_option('dqnpolicy_' + domainString, 'h1_size'):
self.h1_size = cfg.getint('dqnpolicy_' + domainString, 'h1_size')
if cfg.has_option('dqnpolicy_' + domainString, 'h2_size'):
self.h2_size = cfg.getint('dqnpolicy_' + domainString, 'h2_size')
if cfg.has_option('policy_' + domainString, 'save_step'):
self.save_step = cfg.getint('policy_' + domainString, 'save_step')
if cfg.has_option('dqnpolicy_' + domainString, 'importance_sampling'):
self.importance_sampling = cfg.get('dqnpolicy_' + domainString, 'importance_sampling')
if cfg.has_option('dqnpolicy_' + domainString, 'training_frequency'):
self.training_frequency = cfg.getint('dqnpolicy_' + domainString, 'training_frequency')
self.natural_gradient_prev = 0.
"""
self.shuffle = False
if cfg.has_option('dqnpolicy_'+domainString, 'experience_replay'):
self.shuffle = cfg.getboolean('dqnpolicy_'+domainString, 'experience_replay')
if not self.shuffle:
# If we don't use experience replay, we don't need to maintain
# sliding window of experiences with maximum capacity.
# We only need to maintain the data of minibatch_size
self.capacity = self.minibatch_size
"""
self.episode_ave_max_q = []
self.mu_prob = 0. # behavioral policy
os.environ["CUDA_VISIBLE_DEVICES"]=""
# init session
self.sess = tf.Session()
with tf.device("/cpu:0"):
np.random.seed(self.randomseed)
tf.set_random_seed(self.randomseed)
# initialise an replay buffer
if self.replay_type == 'vanilla':
self.episodes[self.domainString] = ReplayBufferEpisode(self.capacity, self.minibatch_size, self.randomseed)
elif self.replay_type == 'prioritized':
self.episodes[self.domainString] = ReplayPrioritisedEpisode(self.capacity, self.minibatch_size, self.randomseed)
#replay_buffer = ReplayBuffer(self.capacity, self.randomseed)
#self.episodes = []
self.samplecount = 0
self.episodecount = 0
# construct the models
self.state_dim = self.n_in
self.summaryaction = SummaryAction.SummaryAction(domainString)
self.action_dim = len(self.summaryaction.action_names)
action_bound = len(self.summaryaction.action_names)
self.stats = [0 for _ in range(self.action_dim)]
self.enac = enac.ENACNetwork(self.sess, self.state_dim, self.action_dim, \
self.critic_lr, self.tau, action_bound, self.architecture, self.h1_size, self.h2_size, self.is_training)
# when all models are defined, init all variables
init_op = tf.global_variables_initializer()
self.sess.run(init_op)
self.loadPolicy(self.in_policy_file)
print 'loaded replay size: ', self.episodes[self.domainString].size()
def __init__(self,
in_policy_file,
out_policy_file,
domainString='CamRestaurants',
is_training=False,
action_names=None):
super(DQNPolicy, self).__init__(domainString, is_training)
tf.reset_default_graph()
self.domainString = domainString
self.domainUtil = FlatOnt.FlatDomainOntology(self.domainString)
self.in_policy_file = in_policy_file
self.out_policy_file = out_policy_file
self.is_training = is_training
self.accum_belief = []
self.domainUtil = FlatOnt.FlatDomainOntology(self.domainString)
self.prev_state_check = None
# pw: Use turn info for predictions
# action vector creation
action_names = [
] # hardcoded to include slots for specific actions (request, confirm, select)
action_names += [
"request(food)", "request(area)", "request(pricerange)",
"confirm(food)", "confirm(area)", "confirm(pricerange)",
"select(food)", "select(area)", "select(pricerange)", "inform",
"inform_byname", "inform_alternatives", "bye", "repeat", "reqmore",
"restart"
]
num_actions = len(action_names)
self.prev_state = None
# parameter settings
if 0: #cfg.has_option('dqnpolicy', 'n_in'): #ic304: this was giving me a weird error, disabled it until i can check it deeper
self.n_in = cfg.getint('dqnpolicy', 'n_in')
else:
self.n_in = self.get_n_in(domainString)
self.learning_rate = 0.001
if cfg.has_option('dqnpolicy', 'learning_rate'):
self.learning_rate = cfg.getfloat('dqnpolicy', 'learning_rate')
self.tau = 0.001
if cfg.has_option('dqnpolicy', 'tau'):
self.tau = cfg.getfloat('dqnpolicy', 'tau')
self.randomseed = 1234
if cfg.has_option('GENERAL', 'seed'):
self.randomseed = cfg.getint('GENERAL', 'seed')
self.gamma = 1.0
if cfg.has_option('dqnpolicy', 'gamma'):
self.gamma = cfg.getfloat('dqnpolicy', 'gamma')
self.regularisation = 'l2'
if cfg.has_option('dqnpolicy', 'regularisation'):
self.regularisation = cfg.get('dqnpolicy', 'regulariser')
self.exploration_type = 'e-greedy' # Boltzman
if cfg.has_option('dqnpolicy', 'exploration_type'):
self.exploration_type = cfg.get('dqnpolicy', 'exploration_type')
self.episodeNum = 1000
if cfg.has_option('dqnpolicy', 'episodeNum'):
self.episodeNum = cfg.getfloat('dqnpolicy', 'episodeNum')
self.maxiter = 5000
if cfg.has_option('dqnpolicy', 'maxiter'):
self.maxiter = cfg.getfloat('dqnpolicy', 'maxiter')
self.epsilon = 1
if cfg.has_option('dqnpolicy', 'epsilon'):
self.epsilon = cfg.getfloat('dqnpolicy', 'epsilon')
self.epsilon_start = 1
if cfg.has_option('dqnpolicy', 'epsilon_start'):
self.epsilon_start = cfg.getfloat('dqnpolicy', 'epsilon_start')
self.epsilon_end = 1
if cfg.has_option('dqnpolicy', 'epsilon_end'):
self.epsilon_end = cfg.getfloat('dqnpolicy', 'epsilon_end')
self.save_step = 100
if cfg.has_option('policy', 'save_step'):
self.save_step = cfg.getint('policy', 'save_step')
self.priorProbStart = 1.0
if cfg.has_option('dqnpolicy', 'prior_sample_prob_start'):
self.priorProbStart = cfg.getfloat('dqnpolicy',
'prior_sample_prob_start')
self.priorProbEnd = 0.1
if cfg.has_option('dqnpolicy', 'prior_sample_prob_end'):
self.priorProbEnd = cfg.getfloat('dqnpolicy',
'prior_sample_prob_end')
self.policyfeatures = []
if cfg.has_option('dqnpolicy', 'features'):
logger.info('Features: ' + str(cfg.get('dqnpolicy', 'features')))
self.policyfeatures = json.loads(cfg.get('dqnpolicy', 'features'))
self.max_k = 5
if cfg.has_option('dqnpolicy', 'max_k'):
self.max_k = cfg.getint('dqnpolicy', 'max_k')
self.learning_algorithm = 'drl'
if cfg.has_option('dqnpolicy', 'learning_algorithm'):
self.learning_algorithm = cfg.get('dqnpolicy',
'learning_algorithm')
logger.info('Learning algorithm: ' + self.learning_algorithm)
self.minibatch_size = 32
if cfg.has_option('dqnpolicy', 'minibatch_size'):
self.minibatch_size = cfg.getint('dqnpolicy', 'minibatch_size')
self.capacity = 1000
if cfg.has_option('dqnpolicy', 'capacity'):
self.capacity = cfg.getint('dqnpolicy', 'capacity')
self.replay_type = 'vanilla'
if cfg.has_option('dqnpolicy', 'replay_type'):
self.replay_type = cfg.get('dqnpolicy', 'replay_type')
self.architecture = 'vanilla'
if cfg.has_option('dqnpolicy', 'architecture'):
self.architecture = cfg.get('dqnpolicy', 'architecture')
if self.architecture == 'dip':
self.architecture = 'dip2'
self.q_update = 'single'
if cfg.has_option('dqnpolicy', 'q_update'):
self.q_update = cfg.get('dqnpolicy', 'q_update')
self.h1_size = 130
if cfg.has_option('dqnpolicy', 'h1_size'):
self.h1_size = cfg.getint('dqnpolicy', 'h1_size')
self.h2_size = 130
if cfg.has_option('dqnpolicy', 'h2_size'):
self.h2_size = cfg.getint('dqnpolicy', 'h2_size')
self.training_frequency = 2
if cfg.has_option('dqnpolicy', 'training_frequency'):
self.training_frequency = cfg.getint('dqnpolicy',
'training_frequency')
# domain specific parameter settings (overrides general policy parameter settings)
if cfg.has_option('dqnpolicy_' + domainString, 'n_in'):
self.n_in = cfg.getint('dqnpolicy_' + domainString, 'n_in')
if cfg.has_option('dqnpolicy_' + domainString, 'learning_rate'):
self.learning_rate = cfg.getfloat('dqnpolicy_' + domainString,
'learning_rate')
if cfg.has_option('dqnpolicy_' + domainString, 'tau'):
self.tau = cfg.getfloat('dqnpolicy_' + domainString, 'tau')
if cfg.has_option('dqnpolicy_' + domainString, 'gamma'):
self.gamma = cfg.getfloat('dqnpolicy_' + domainString, 'gamma')
if cfg.has_option('dqnpolicy_' + domainString, 'regularisation'):
self.regularisation = cfg.get('dqnpolicy_' + domainString,
'regulariser')
if cfg.has_option('dqnpolicy_' + domainString, 'exploration_type'):
self.exploration_type = cfg.get('dqnpolicy_' + domainString,
'exploration_type')
if cfg.has_option('dqnpolicy_' + domainString, 'episodeNum'):
self.episodeNum = cfg.getfloat('dqnpolicy_' + domainString,
'episodeNum')
if cfg.has_option('dqnpolicy_' + domainString, 'maxiter'):
self.maxiter = cfg.getfloat('dqnpolicy_' + domainString, 'maxiter')
if cfg.has_option('dqnpolicy_' + domainString, 'epsilon'):
self.epsilon = cfg.getfloat('dqnpolicy_' + domainString, 'epsilon')
if cfg.has_option('dqnpolicy_' + domainString, 'epsilon_start'):
self.epsilon_start = cfg.getfloat('dqnpolicy_' + domainString,
'epsilon_start')
if cfg.has_option('dqnpolicy_' + domainString, 'epsilon_end'):
self.epsilon_end = cfg.getfloat('dqnpolicy_' + domainString,
'epsilon_end')
if cfg.has_option('policy_' + domainString, 'save_step'):
self.save_step = cfg.getint('policy_' + domainString, 'save_step')
if cfg.has_option('dqnpolicy_' + domainString,
'prior_sample_prob_start'):
self.priorProbStart = cfg.getfloat('dqnpolicy_' + domainString,
'prior_sample_prob_start')
if cfg.has_option('dqnpolicy_' + domainString,
'prior_sample_prob_end'):
self.priorProbEnd = cfg.getfloat('dqnpolicy_' + domainString,
'prior_sample_prob_end')
if cfg.has_option('dqnpolicy_' + domainString, 'features'):
logger.info('Features: ' +
str(cfg.get('dqnpolicy_' + domainString, 'features')))
self.policyfeatures = json.loads(
cfg.get('dqnpolicy_' + domainString, 'features'))
if cfg.has_option('dqnpolicy_' + domainString, 'max_k'):
self.max_k = cfg.getint('dqnpolicy_' + domainString, 'max_k')
if cfg.has_option('dqnpolicy_' + domainString, 'learning_algorithm'):
self.learning_algorithm = cfg.get('dqnpolicy_' + domainString,
'learning_algorithm')
logger.info('Learning algorithm: ' + self.learning_algorithm)
if cfg.has_option('dqnpolicy_' + domainString, 'minibatch_size'):
self.minibatch_size = cfg.getint('dqnpolicy_' + domainString,
'minibatch_size')
if cfg.has_option('dqnpolicy_' + domainString, 'capacity'):
self.capacity = cfg.getint('dqnpolicy_' + domainString, 'capacity')
if cfg.has_option('dqnpolicy_' + domainString, 'replay_type'):
self.replay_type = cfg.get('dqnpolicy_' + domainString,
'replay_type')
if cfg.has_option('dqnpolicy_' + domainString, 'architecture'):
self.architecture = cfg.get('dqnpolicy_' + domainString,
'architecture')
if cfg.has_option('dqnpolicy_' + domainString, 'q_update'):
self.q_update = cfg.get('dqnpolicy_' + domainString, 'q_update')
if cfg.has_option('dqnpolicy_' + domainString, 'h1_size'):
self.h1_size = cfg.getint('dqnpolicy_' + domainString, 'h1_size')
if cfg.has_option('dqnpolicy_' + domainString, 'h2_size'):
self.h2_size = cfg.getint('dqnpolicy_' + domainString, 'h2_size')
if cfg.has_option('dqnpolicy_' + domainString, 'training_frequency'):
self.training_frequency = cfg.getint('dqnpolicy_' + domainString,
'training_frequency')
"""
self.shuffle = False
if cfg.has_option('dqnpolicy_'+domainString, 'experience_replay'):
self.shuffle = cfg.getboolean('dqnpolicy_'+domainString, 'experience_replay')
if not self.shuffle:
# If we don't use experience replay, we don't need to maintain
# sliding window of experiences with maximum capacity.
# We only need to maintain the data of minibatch_size
self.capacity = self.minibatch_size
"""
self.episode_ave_max_q = []
os.environ["CUDA_VISIBLE_DEVICES"] = ""
policytype = 'dqn'
self.dropout_rate = 0.
if cfg.has_option('dqnpolicy', 'dropout_rate'):
self.dropout_rate = cfg.getfloat('dqnpolicy', 'dropout_rate')
if cfg.has_option('policy', 'policytype'):
policytype = cfg.get('policy', 'policytype')
if policytype != 'feudal':
# init session
self.sess = tf.Session()
with tf.device("/cpu:0"):
np.random.seed(self.randomseed)
tf.set_random_seed(self.randomseed)
# initialise an replay buffer
if self.replay_type == 'vanilla':
self.episodes[self.domainString] = ReplayBuffer(
self.capacity, self.minibatch_size, self.randomseed)
elif self.replay_type == 'prioritized':
self.episodes[self.domainString] = ReplayPrioritised(
self.capacity, self.minibatch_size, self.randomseed)
self.samplecount = 0
self.episodecount = 0
# construct the models
self.state_dim = self.n_in
if self.architecture == 'dip2':
self.state_dim = 89
self.summaryaction = SummaryAction.SummaryAction(domainString)
if action_names is None:
self.action_names = self.summaryaction.action_names
else:
self.action_names = action_names
self.action_dim = len(self.action_names)
action_bound = len(self.action_names)
self.stats = [0 for _ in range(self.action_dim)]
self.dqn = dqn.DeepQNetwork(self.sess, self.state_dim, self.action_dim, \
self.learning_rate, self.tau, action_bound, self.minibatch_size,
self.architecture, self.h1_size,
self.h2_size, dropout_rate=self.dropout_rate)
# when all models are defined, init all variables
init_op = tf.global_variables_initializer()
self.sess.run(init_op)
self.loadPolicy(self.in_policy_file)
print 'loaded replay size: ', self.episodes[
self.domainString].size()
self.dqn.update_target_network()
def __init__(self,
in_policy_file,
out_policy_file,
domainString='CamRestaurants',
is_training=False):
super(FeudalPolicy, self).__init__(domainString, is_training)
self.domainString = domainString
self.domainUtil = FlatOnt.FlatDomainOntology(self.domainString)
self.in_policy_file = in_policy_file
self.out_policy_file = out_policy_file
self.is_training = is_training
self.prev_state_check = None
#feudalRL variables
self.prev_sub_policy = None
self.prev_master_act = None
self.prev_master_belief = None
self.prev_child_act = None
self.prev_child_belief = None
self.action_freq = np.zeros(len(self.actions.action_names))
self.master_dec_count = np.array([0., 0.])
self.gi_dec_inrow = 0
self.features = 'dip'
if cfg.has_option('feudalpolicy', 'features'):
self.features = cfg.get('feudalpolicy', 'features')
self.si_policy_type = 'dqn'
if cfg.has_option('feudalpolicy', 'si_policy_type'):
self.si_policy_type = cfg.get('feudalpolicy', 'si_policy_type')
self.sd_policy_type = 'dqn'
if cfg.has_option('feudalpolicy', 'sd_policy_type'):
self.sd_policy_type = cfg.get('feudalpolicy', 'sd_policy_type')
self.master_policy_type = self.si_policy_type
if cfg.has_option('feudalpolicy', 'master_policy_type'):
self.master_policy_type = cfg.get('feudalpolicy',
'master_policy_type')
self.sample_master = False
if cfg.has_option('feudalpolicy', 'sample_master'):
self.sample_master = cfg.getboolean('feudalpolicy',
'sample_master')
self.correct_master = False
if cfg.has_option('feudalpolicy', 'correct_master'):
self.correct_master = cfg.getboolean('feudalpolicy',
'correct_master')
self.use_bye = False
if cfg.has_option('feudalpolicy', 'use_bye'):
self.use_bye = cfg.getboolean('feudalpolicy', 'use_bye')
self.reqmore_in_si = True
if cfg.has_option('feudalpolicy', 'reqmore_in_si'):
self.reqmore_in_si = cfg.getboolean('feudalpolicy',
'reqmore_in_si')
self.correction_factor = 0
if cfg.has_option('feudalpolicy', 'correction_factor'):
self.correction_factor = cfg.getfloat('feudalpolicy',
'correction_factor')
self.actfreq_ds = False
if cfg.has_option('feudalpolicy', 'actfreq_ds'):
self.actfreq_ds = cfg.getboolean('feudalpolicy', 'actfreq_ds')
# parameter settings
self.randomseed = 1234
if cfg.has_option('GENERAL', 'seed'):
self.randomseed = cfg.getint('GENERAL', 'seed')
# Create the feudal structure (including feudal masks)
self.summaryaction = SummaryAction.SummaryAction(domainString)
self.full_action_list = self.summaryaction.action_names
self.master_actions = ['give_info', 'request_info', 'pass']
self.slot_independent_actions = [
"inform", "inform_byname", "inform_alternatives"
]
if self.reqmore_in_si:
self.slot_independent_actions.append("reqmore")
if self.use_bye:
self.slot_independent_actions.append('bye')
self.slot_independent_actions.append('pass')
self.slot_specific_actions = ["request", "confirm", "select"]
#if self.reqmore_in_sd is True:
# self.slot_specific_actions.append("reqmore")
self.slot_specific_actions.append('pass')
self.master_freq = np.zeros(len(self.master_actions))
self.si_freq = np.zeros(len(self.slot_independent_actions))
self.sd_freq = np.zeros(len(self.slot_specific_actions))
# master policy
if self.master_policy_type == 'acer':
self.master_policy = FeudalACERPolicy(
self._modify_policyfile('master', in_policy_file),
self._modify_policyfile('master', out_policy_file),
domainString=self.domainString,
is_training=self.is_training,
action_names=['give_info', 'request_info', 'pass'],
slot='si'
) # pass is always masked, but its needed for implementation
elif self.master_policy_type == 'enac':
self.master_policy = FeudalENACPolicy(
self._modify_policyfile('master', in_policy_file),
self._modify_policyfile('master', out_policy_file),
domainString=self.domainString,
is_training=self.is_training,
action_names=['give_info', 'request_info', 'pass'],
slot='si'
) # pass is always masked, but its needed for implementation
elif self.master_policy_type == 'bbqn':
self.master_policy = FeudalBBQNPolicy(
self._modify_policyfile('master', in_policy_file),
self._modify_policyfile('master', out_policy_file),
domainString=self.domainString,
is_training=self.is_training,
action_names=['give_info', 'request_info', 'pass'],
slot='si'
) # pass is always masked, but its needed for implementation
else:
self.master_policy = FeudalDQNPolicy(
self._modify_policyfile('master', in_policy_file),
self._modify_policyfile('master', out_policy_file),
domainString=self.domainString,
is_training=self.is_training,
action_names=['give_info', 'request_info', 'pass'],
slot='si'
) # pass is always masked, but its needed for implementation
# si policy
if self.si_policy_type == 'acer':
self.give_info_policy = FeudalACERPolicy(
self._modify_policyfile('gi', in_policy_file),
self._modify_policyfile('gi', out_policy_file),
domainString=self.domainString,
is_training=self.is_training,
action_names=self.slot_independent_actions,
slot='si')
elif self.si_policy_type == 'enac':
self.give_info_policy = FeudalENACPolicy(
self._modify_policyfile('gi', in_policy_file),
self._modify_policyfile('gi', out_policy_file),
domainString=self.domainString,
is_training=self.is_training,
action_names=self.slot_independent_actions,
slot='si')
elif self.si_policy_type == 'bbqn':
self.give_info_policy = FeudalBBQNPolicy(
self._modify_policyfile('gi', in_policy_file),
self._modify_policyfile('gi', out_policy_file),
domainString=self.domainString,
is_training=self.is_training,
action_names=self.slot_independent_actions,
slot='si')
else:
self.give_info_policy = FeudalDQNPolicy(
self._modify_policyfile('gi', in_policy_file),
self._modify_policyfile('gi', out_policy_file),
domainString=self.domainString,
is_training=self.is_training,
action_names=self.slot_independent_actions,
slot='si')
# sd policies
if self.sd_policy_type == 'acer':
self.request_info_policy = FeudalACERPolicy(
self._modify_policyfile('ri', in_policy_file),
self._modify_policyfile('ri', out_policy_file),
domainString=self.domainString,
is_training=self.is_training,
action_names=self.slot_specific_actions,
slot='sd')
elif self.sd_policy_type == 'bbqn':
self.request_info_policy = FeudalBBQNPolicy(
self._modify_policyfile('ri', in_policy_file),
self._modify_policyfile('ri', out_policy_file),
domainString=self.domainString,
is_training=self.is_training,
action_names=self.slot_specific_actions,
slot='sd')
else:
self.request_info_policy = FeudalDQNPolicy(
self._modify_policyfile('ri', in_policy_file),
self._modify_policyfile('ri', out_policy_file),
domainString=self.domainString,
is_training=self.is_training,
action_names=self.slot_specific_actions,
slot='sd')
def __init__(self,
in_policy_file,
out_policy_file,
domainString='CamRestaurants',
is_training=False,
action_names=None):
super(MORLPolicy, self).__init__(domainString, is_training)
self.domainString = domainString
self.domainUtil = FlatOnt.FlatDomainOntology(self.domainString)
self.in_policy_file = in_policy_file
self.out_policy_file = out_policy_file
self.is_training = is_training
self.accum_belief = []
self.domainUtil = FlatOnt.FlatDomainOntology(self.domainString)
self.prev_state_check = None
# parameter settings
if 0: # cfg.has_option('morlpolicy', 'n_in'): #ic304: this was giving me a weird error, disabled it until i can check it deeper
self.n_in = cfg.getint('morlpolicy', 'n_in')
else:
self.n_in = self.get_n_in(domainString)
self.n_rew = 1
if cfg.has_option('morlpolicy', 'n_rew'):
self.n_rew = cfg.getint('morlpolicy', 'n_rew')
self.lr = 0.001
if cfg.has_option('morlpolicy', 'learning_rate'):
self.lr = cfg.getfloat('morlpolicy', 'learning_rate')
self.epsilon = 0.5
if cfg.has_option('morlpolicy', 'epsilon'):
self.epsilon = cfg.getfloat('morlpolicy', 'epsilon')
self.epsilon_decay = True
if cfg.has_option('morlpolicy', 'epsilon_decay'):
self.epsilon_decay = cfg.getboolean('morlpolicy', 'epsilon_decay')
self.randomseed = 1234
if cfg.has_option('GENERAL', 'seed'):
self.randomseed = cfg.getint('GENERAL', 'seed')
self.gamma = 1.0
if cfg.has_option('morlpolicy', 'gamma'):
self.gamma = cfg.getfloat('morlpolicy', 'gamma')
self.weight_num = 32
if cfg.has_option('morlpolicy', 'weight_num'):
self.weight_num = cfg.getint('morlpolicy', 'weight_num')
self.episode_num = 1000
if cfg.has_option('morlpolicy', 'episode_num'):
self.episode_num = cfg.getfloat('morlpolicy', 'episode_num')
self.optimizer = "Adam"
if cfg.has_option('morlpolicy', 'optimizer'):
self.optimizer = cfg.get('morlpolicy', 'optimizer')
self.save_step = 100
if cfg.has_option('policy', 'save_step'):
self.save_step = cfg.getint('policy', 'save_step')
self.update_freq = 50
if cfg.has_option('morlpolicy', 'update_freq'):
self.update_freq = cfg.getint('morlpolicy', 'update_freq')
self.policyfeatures = []
if cfg.has_option('morlpolicy', 'features'):
logger.info('Features: ' + str(cfg.get('morlpolicy', 'features')))
self.policyfeatures = json.loads(cfg.get('morlpolicy', 'features'))
self.algorithm = 'naive'
if cfg.has_option('morlpolicy', 'algorithm'):
self.algorithm = cfg.get('morlpolicy', 'algorithm')
logger.info('Learning algorithm: ' + self.algorithm)
self.batch_size = 32
if cfg.has_option('morlpolicy', 'batch_size'):
self.batch_size = cfg.getint('morlpolicy', 'batch_size')
self.mem_size = 1000
if cfg.has_option('morlpolicy', 'mem_size'):
self.mem_size = cfg.getint('morlpolicy', 'mem_size')
self.training_freq = 1
if cfg.has_option('morlpolicy', 'training_freq'):
self.training_freq = cfg.getint('morlpolicy', 'training_freq')
# set beta for envelope algorithm
self.beta = 0.1
if cfg.has_option('morlpolicy', 'beta'):
self.beta = cfg.getfloat('morlpolicy', 'beta')
self.beta_init = self.beta
self.beta_uplim = 1.00
self.tau = 1000.
self.beta_expbase = float(
np.power(self.tau * (self.beta_uplim - self.beta),
1. / (self.episode_num + 1)))
self.beta_delta = self.beta_expbase / self.tau
self.beta -= self.beta_delta
# using homotopy method for optimization
self.homotopy = False
if cfg.has_option('morlpolicy', 'homotopy'):
self.homotopy = cfg.getboolean('morlpolicy', 'homotopy')
self.epsilon_delta = (self.epsilon - 0.05) / self.episode_num
self.episodecount = 0
# construct the models
self.state_dim = self.n_in
self.summaryaction = SummaryAction.SummaryAction(domainString)
if action_names is None:
self.action_names = self.summaryaction.action_names
else:
self.action_names = action_names
self.action_dim = len(self.action_names)
self.stats = [0 for _ in range(self.action_dim)]
self.reward_dim = self.n_rew
model = None
if self.algorithm == 'naive':
model = naive.NaiveLinearCQN(self.state_dim, self.action_dim,
self.reward_dim)
elif self.algorithm == 'envelope':
model = envelope.EnvelopeLinearCQN(self.state_dim, self.action_dim,
self.reward_dim)
self.model_ = model
self.model = copy.deepcopy(model)
# initialize memory
self.trans_mem = deque()
self.trans = namedtuple('trans',
['s', 'a', 's_', 'r', 'd', 'ms', 'ms_'])
self.priority_mem = deque()
self.mem_last_state = None
self.mem_last_action = None
self.mem_last_mask = None
self.mem_cur_state = None
self.mem_cur_action = None
self.mem_cur_mask = None
if self.optimizer == 'Adam':
self.optimizer = optim.Adam(self.model_.parameters(), lr=self.lr)
elif self.optimizer == 'RMSprop':
self.optimizer = optim.RMSprop(self.model_.parameters(),
lr=self.lr)
try:
self.loadPolicy(self.in_policy_file)
except:
logger.info("No previous model found...")
self.w_kept = None
self.update_count = 0
if self.is_training:
self.model_.train()
if use_cuda:
self.model.cuda()
self.model_.cuda()
self.monitor = None
def __init__(self,
in_policy_file,
out_policy_file,
domainString='CamRestaurants',
is_training=False,
action_names=None,
slot=None):
super(FeudalACERPolicy, self).__init__(in_policy_file, out_policy_file,
domainString, is_training)
tf.reset_default_graph()
self.in_policy_file = in_policy_file
self.out_policy_file = out_policy_file
self.is_training = is_training
self.accum_belief = []
self.prev_state_check = None
self.domainString = domainString
self.domainUtil = FlatOnt.FlatDomainOntology(self.domainString)
self.features = 'dip'
self.sd_enc_size = 80
self.si_enc_size = 40
self.dropout_rate = 0.
if cfg.has_option('feudalpolicy', 'features'):
self.features = cfg.get('feudalpolicy', 'features')
if cfg.has_option('feudalpolicy', 'sd_enc_size'):
self.sd_enc_size = cfg.getint('feudalpolicy', 'sd_enc_size')
if cfg.has_option('feudalpolicy', 'si_enc_size'):
self.si_enc_size = cfg.getint('feudalpolicy', 'si_enc_size')
if cfg.has_option('dqnpolicy', 'dropout_rate') and self.is_training:
self.dropout_rate = cfg.getfloat('feudalpolicy', 'dropout_rate')
if cfg.has_option('dqnpolicy', 'dropout_rate') and self.is_training:
self.dropout_rate = cfg.getfloat('feudalpolicy', 'dropout_rate')
self.actfreq_ds = False
if cfg.has_option('feudalpolicy', 'actfreq_ds'):
self.actfreq_ds = cfg.getboolean('feudalpolicy', 'actfreq_ds')
# init session
self.sess = tf.Session()
with tf.device("/cpu:0"):
np.random.seed(self.randomseed)
tf.set_random_seed(self.randomseed)
# initialise an replay buffer
if self.replay_type == 'vanilla':
self.episodes[self.domainString] = ReplayBufferEpisode(
self.capacity, self.minibatch_size, self.randomseed)
elif self.replay_type == 'prioritized':
self.episodes[self.domainString] = ReplayPrioritisedEpisode(
self.capacity, self.minibatch_size, self.randomseed)
#replay_buffer = ReplayBuffer(self.capacity, self.randomseed)
#self.episodes = []
self.samplecount = 0
self.episodecount = 0
# construct the models
self.state_dim = 89 # current DIP state dim
self.summaryaction = policy.SummaryAction.SummaryAction(
domainString)
self.action_names = action_names
self.action_dim = len(self.action_names)
action_bound = len(self.action_names)
self.stats = [0 for _ in range(self.action_dim)]
self.global_mu = [0. for _ in range(self.action_dim)]
if self.features == 'dip':
if self.actfreq_ds:
if self.domainString == 'CamRestaurants':
self.state_dim += 9 #16
elif self.domainString == 'SFRestaurants':
self.state_dim += 9 #25
elif self.domainString == 'Laptops11':
self.state_dim += 9 #40
self.acer = acer.ACERNetwork(self.sess, self.state_dim,
self.action_dim, self.critic_lr,
self.delta, self.c, self.alpha,
self.h1_size, self.h2_size,
self.is_training)
elif self.features == 'learned' or self.features == 'rnn':
si_state_dim = 72
if self.actfreq_ds:
if self.domainString == 'CamRestaurants':
si_state_dim += 9 #16
elif self.domainString == 'SFRestaurants':
si_state_dim += 9 #25
elif self.domainString == 'Laptops11':
si_state_dim += 9 #40
if self.domainString == 'CamRestaurants':
sd_state_dim = 158 #94
elif self.domainString == 'SFRestaurants':
sd_state_dim = 158
elif self.domainString == 'Laptops11':
sd_state_dim = 158 #13
else:
logger.error(
'Domain {} not implemented in feudal-DQN yet'
) # just find out the size of sd_state_dim for the new domain
if 0: #self.features == 'rnn':
self.acer = acer.RNNACERNetwork(self.sess,
si_state_dim,
sd_state_dim,
self.action_dim,
self.critic_lr,
self.delta,
self.c,
self.alpha,
self.h1_size,
self.h2_size,
self.is_training,
sd_enc_size=25,
si_enc_size=25,
dropout_rate=0.,
tn='normal',
slot='si')
else:
self.state_dim = si_state_dim + sd_state_dim
self.acer = acer.ACERNetwork(self.sess, self.state_dim,
self.action_dim,
self.critic_lr, self.delta,
self.c, self.alpha,
self.h1_size, self.h2_size,
self.is_training)
else:
logger.error('features "{}" not implemented'.format(
self.features))
# when all models are defined, init all variables
init_op = tf.global_variables_initializer()
self.sess.run(init_op)
self.loadPolicy(self.in_policy_file)
print 'loaded replay size: ', self.episodes[
self.domainString].size()
def __init__(self,
in_policy_file,
out_policy_file,
domainString='CamRestaurants',
is_training=False,
action_names=None,
slot=None):
super(FeudalDQNPolicy, self).__init__(in_policy_file, out_policy_file,
domainString, is_training)
tf.reset_default_graph()
self.domainString = domainString
self.domainUtil = FlatOnt.FlatDomainOntology(self.domainString)
self.in_policy_file = in_policy_file
self.out_policy_file = out_policy_file
self.is_training = is_training
self.accum_belief = []
self.slot = slot
self.features = 'dip'
if cfg.has_option('feudalpolicy', 'features'):
self.features = cfg.get('feudalpolicy', 'features')
self.actfreq_ds = False
if cfg.has_option('feudalpolicy', 'actfreq_ds'):
self.actfreq_ds = cfg.getboolean('feudalpolicy', 'actfreq_ds')
self.domainUtil = FlatOnt.FlatDomainOntology(self.domainString)
self.prev_state_check = None
self.max_k = 5
if cfg.has_option('dqnpolicy', 'max_k'):
self.max_k = cfg.getint('dqnpolicy', 'max_k')
self.capacity *= 4 # capacity for episode methods, multiply it to adjust to turn based methods
# init session
self.sess = tf.Session()
with tf.device("/cpu:0"):
np.random.seed(self.randomseed)
tf.set_random_seed(self.randomseed)
# initialise a replay buffer
if self.replay_type == 'vanilla':
self.episodes[self.domainString] = ReplayBuffer(
self.capacity, self.minibatch_size, self.randomseed)
elif self.replay_type == 'prioritized':
self.episodes[self.domainString] = ReplayPrioritised(
self.capacity, self.minibatch_size, self.randomseed)
self.samplecount = 0
self.episodecount = 0
# construct the models
self.state_dim = 89 # current DIP state dim
self.summaryaction = policy.SummaryAction.SummaryAction(
domainString)
self.action_names = action_names
self.action_dim = len(self.action_names)
action_bound = len(self.action_names)
self.stats = [0 for _ in range(self.action_dim)]
if self.features == 'learned' or self.features == 'rnn':
si_state_dim = 72
if self.actfreq_ds:
if self.domainString == 'CamRestaurants':
si_state_dim += 9 #16
elif self.domainString == 'SFRestaurants':
si_state_dim += 9 #25
elif self.domainString == 'Laptops11':
si_state_dim += 9 #40
if self.domainString == 'CamRestaurants':
sd_state_dim = 158 #94
elif self.domainString == 'SFRestaurants':
sd_state_dim = 158
elif self.domainString == 'Laptops11':
sd_state_dim = 158 #13
else:
logger.error(
'Domain {} not implemented in feudal-DQN yet'
) # just find out the size of sd_state_dim for the new domain
self.sd_enc_size = 50
self.si_enc_size = 25
self.dropout_rate = 0.
if cfg.has_option('feudalpolicy', 'sd_enc_size'):
self.sd_enc_size = cfg.getint('feudalpolicy',
'sd_enc_size')
if cfg.has_option('feudalpolicy', 'si_enc_size'):
self.si_enc_size = cfg.getint('feudalpolicy',
'si_enc_size')
if cfg.has_option('dqnpolicy',
'dropout_rate') and self.is_training:
self.dropout_rate = cfg.getfloat('feudalpolicy',
'dropout_rate')
if cfg.has_option('dqnpolicy',
'dropout_rate') and self.is_training:
self.dropout_rate = cfg.getfloat('feudalpolicy',
'dropout_rate')
self.state_dim = si_state_dim + sd_state_dim
if self.features == 'learned':
self.dqn = dqn.NNFDeepQNetwork(
self.sess,
si_state_dim,
sd_state_dim,
self.action_dim,
self.learning_rate,
self.tau,
action_bound,
self.minibatch_size,
self.architecture,
self.h1_size,
self.h2_size,
sd_enc_size=self.sd_enc_size,
si_enc_size=self.si_enc_size,
dropout_rate=self.dropout_rate)
elif self.features == 'rnn':
self.dqn = dqn.RNNFDeepQNetwork(
self.sess,
si_state_dim,
sd_state_dim,
self.action_dim,
self.learning_rate,
self.tau,
action_bound,
self.minibatch_size,
self.architecture,
self.h1_size,
self.h2_size,
sd_enc_size=self.sd_enc_size,
si_enc_size=self.si_enc_size,
dropout_rate=self.dropout_rate,
slot=self.slot)
else: # self.features = 'dip'
if self.actfreq_ds:
if self.domainString == 'CamRestaurants':
self.state_dim += 9 #16
elif self.domainString == 'SFRestaurants':
self.state_dim += 9 #25
elif self.domainString == 'Laptops11':
self.state_dim += 9 #40
self.dqn = dqn.DeepQNetwork(self.sess,
self.state_dim,
self.action_dim,
self.learning_rate,
self.tau,
action_bound,
self.minibatch_size,
self.architecture,
self.h1_size,
self.h2_size,
dropout_rate=self.dropout_rate)
# when all models are defined, init all variables (this might to be sent to the main policy too)
init_op = tf.global_variables_initializer()
self.sess.run(init_op)
self.loadPolicy(self.in_policy_file)
print 'loaded replay size: ', self.episodes[
self.domainString].size()
self.dqn.update_target_network()
def __init__(self,
in_policy_file,
out_policy_file,
domainString='CamRestaurants',
is_training=False):
super(BootstrappedDQNPolicy, self).__init__(domainString, is_training)
self.domainString = domainString
self.domainUtil = FlatOnt.FlatDomainOntology(self.domainString)
self.in_policy_file = in_policy_file
self.out_policy_file = out_policy_file
self.is_training = is_training
self.accum_belief = []
self.prev_state_check = None
# parameter settings
self.n_in = 260
if cfg.has_option('dqnpolicy_' + domainString, 'n_in'):
self.n_in = cfg.getint('dqnpolicy_' + domainString, 'n_in')
self.actor_lr = 0.0001
if cfg.has_option('dqnpolicy_' + domainString, 'actor_lr'):
self.actor_lr = cfg.getfloat('dqnpolicy_' + domainString,
'actor_lr')
self.critic_lr = 0.001
if cfg.has_option('dqnpolicy_' + domainString, 'critic_lr'):
self.critic_lr = cfg.getfloat('dqnpolicy_' + domainString,
'critic_lr')
self.tau = 0.001
if cfg.has_option('dqnpolicy_' + domainString, 'tau'):
self.tau = cfg.getfloat('dqnpolicy_' + domainString, 'tau')
self.randomseed = 1234
if cfg.has_option('GENERAL', 'seed'):
self.randomseed = cfg.getint('GENERAL', 'seed')
self.gamma = 1.0
if cfg.has_option('dqnpolicy_' + domainString, 'gamma'):
self.gamma = cfg.getfloat('dqnpolicy_' + domainString, 'gamma')
self.regularisation = 'l2'
if cfg.has_option('dqnpolicy_' + domainString, 'regularisation'):
self.regularisation = cfg.get('dqnpolicy_' + domainString,
'regulariser')
self.learning_rate = 0.001
if cfg.has_option('dqnpolicy_' + domainString, 'learning_rate'):
self.learning_rate = cfg.getfloat('dqnpolicy_' + domainString,
'learning_rate')
self.exploration_type = 'e-greedy' # Boltzman
if cfg.has_option('dqnpolicy_' + domainString, 'exploration_type'):
self.exploration_type = cfg.get('dqnpolicy_' + domainString,
'exploration_type')
self.episodeNum = 1000
if cfg.has_option('dqnpolicy_' + domainString, 'episodeNum'):
self.episodeNum = cfg.getfloat('dqnpolicy_' + domainString,
'episodeNum')
self.maxiter = 5000
if cfg.has_option('dqnpolicy_' + domainString, 'maxiter'):
self.maxiter = cfg.getfloat('dqnpolicy_' + domainString, 'maxiter')
self.epsilon = 1
if cfg.has_option('dqnpolicy_' + domainString, 'epsilon'):
self.epsilon = cfg.getfloat('dqnpolicy_' + domainString, 'epsilon')
self.epsilon_start = 1
if cfg.has_option('dqnpolicy_' + domainString, 'epsilon_start'):
self.epsilon_start = cfg.getfloat('dqnpolicy_' + domainString,
'epsilon_start')
self.epsilon_end = 1
if cfg.has_option('dqnpolicy_' + domainString, 'epsilon_end'):
self.epsilon_end = cfg.getfloat('dqnpolicy_' + domainString,
'epsilon_end')
self.priorProbStart = 1.0
if cfg.has_option('dqnpolicy_' + domainString,
'prior_sample_prob_start'):
self.priorProbStart = cfg.getfloat('dqnpolicy_' + domainString,
'prior_sample_prob_start')
self.save_step = 100
if cfg.has_option('policy_' + domainString, 'save_step'):
self.save_step = cfg.getint('policy_' + domainString, 'save_step')
self.priorProbEnd = 0.1
if cfg.has_option('dqnpolicy_' + domainString,
'prior_sample_prob_end'):
self.priorProbEnd = cfg.getfloat('dqnpolicy_' + domainString,
'prior_sample_prob_end')
self.policyfeatures = []
if cfg.has_option('dqnpolicy_' + domainString, 'features'):
logger.info('Features: ' +
str(cfg.get('dqnpolicy_' + domainString, 'features')))
self.policyfeatures = json.loads(
cfg.get('dqnpolicy_' + domainString, 'features'))
self.max_k = 5
if cfg.has_option('dqnpolicy_' + domainString, 'max_k'):
self.max_k = cfg.getint('dqnpolicy_' + domainString, 'max_k')
self.learning_algorithm = 'drl'
if cfg.has_option('dqnpolicy_' + domainString, 'learning_algorithm'):
self.learning_algorithm = cfg.get('dqnpolicy_' + domainString,
'learning_algorithm')
logger.info('Learning algorithm: ' + self.learning_algorithm)
self.minibatch_size = 32
if cfg.has_option('dqnpolicy_' + domainString, 'minibatch_size'):
self.minibatch_size = cfg.getint('dqnpolicy_' + domainString,
'minibatch_size')
self.capacity = 1000 # max(self.minibatch_size, 2000)
if cfg.has_option('dqnpolicy_' + domainString, 'capacity'):
self.capacity = max(
cfg.getint('dqnpolicy_' + domainString, 'capacity'), 2000)
self.replay_type = 'vanilla'
if cfg.has_option('dqnpolicy_' + domainString, 'replay_type'):
self.replay_type = cfg.get('dqnpolicy_' + domainString,
'replay_type')
self.architecture = 'vanilla'
if cfg.has_option('dqnpolicy_' + domainString, 'architecture'):
self.architecture = cfg.get('dqnpolicy_' + domainString,
'architecture')
self.q_update = 'single'
if cfg.has_option('dqnpolicy_' + domainString, 'q_update'):
self.q_update = cfg.get('dqnpolicy_' + domainString, 'q_update')
self.h1_size = 130
if cfg.has_option('dqnpolicy_' + domainString, 'h1_size'):
self.h1_size = cfg.getint('dqnpolicy_' + domainString, 'h1_size')
self.h2_size = 130
if cfg.has_option('dqnpolicy_' + domainString, 'h2_size'):
self.h2_size = cfg.getint('dqnpolicy_' + domainString, 'h2_size')
self.no_heads = 3
if cfg.has_option('dqnpolicy_' + domainString, 'no_head'):
self.no_heads = cfg.getint('dqnpolicy_' + domainString, 'no_head')
self.episode_ave_max_q = []
os.environ["CUDA_VISIBLE_DEVICES"] = ""
# initialize head
self.head = None
# init session
self.sess = tf.Session()
with tf.device("/cpu:0"):
np.random.seed(self.randomseed)
tf.set_random_seed(self.randomseed)
# initialise an replay buffer
if self.replay_type == 'vanilla':
self.episodes[self.domainString] = ReplayBuffer(
self.capacity, self.minibatch_size, self.randomseed)
elif self.replay_type == 'prioritized':
self.episodes[self.domainString] = ReplayPrioritised(
self.capacity, self.minibatch_size, self.randomseed)
# replay_buffer = ReplayBuffer(self.capacity, self.randomseed)
# self.episodes = []
self.samplecount = 0
self.episodecount = 0
# construct the models
self.state_dim = self.n_in
self.summaryaction = SummaryAction.SummaryAction(domainString)
self.action_dim = len(self.summaryaction.action_names)
action_bound = len(self.summaryaction.action_names)
self.stats = [0 for _ in range(self.action_dim)]
self.dqn = dqn.DeepQNetwork(self.sess, self.state_dim, self.action_dim, \
self.critic_lr, self.tau, action_bound, self.architecture, self.h1_size,
self.h2_size, self.no_heads, self.minibatch_size)
# when all models are defined, init all variables
init_op = tf.global_variables_initializer()
self.sess.run(init_op)
self.loadPolicy(self.in_policy_file)
print 'loaded replay size: ', self.episodes[
self.domainString].size()
for head in range(self.no_heads):
self.dqn.update_target_network(head)
def __init__(self,
in_policy_file,
out_policy_file,
domainString='CamRestaurants',
is_training=False):
super(ACERPolicy, self).__init__(domainString, is_training)
tf.reset_default_graph()
self.in_policy_file = in_policy_file
self.out_policy_file = out_policy_file
self.is_training = is_training
self.accum_belief = []
self.prev_state_check = None
self.domainString = domainString
self.domainUtil = FlatOnt.FlatDomainOntology(self.domainString)
#improvement==================================
self.intrinsic_reward_method = None
if cfg.has_option('scme', 'method'):
self.intrinsic_reward_method = cfg.get('scme', 'method')
#improvement==================================
# parameter settings
if 0: # cfg.has_option('dqnpolicy', 'n_in'): #ic304: this was giving me a weird error, disabled it until i can check it deeper
self.n_in = cfg.getint('dqnpolicy', 'n_in')
else:
self.n_in = self.get_n_in(domainString)
self.actor_lr = 0.0001
if cfg.has_option('dqnpolicy', 'actor_lr'):
self.actor_lr = cfg.getfloat('dqnpolicy', 'actor_lr')
self.critic_lr = 0.001
if cfg.has_option('dqnpolicy', 'critic_lr'):
self.critic_lr = cfg.getfloat('dqnpolicy', 'critic_lr')
self.delta = 1.
if cfg.has_option('dqnpolicy', 'delta'):
self.delta = cfg.getfloat('dqnpolicy', 'delta')
self.alpha = 0.99
if cfg.has_option('dqnpolicy', 'beta'):
self.alpha = cfg.getfloat('dqnpolicy', 'beta')
self.c = 10.
if cfg.has_option('dqnpolicy', 'is_threshold'):
self.c = cfg.getfloat('dqnpolicy', 'is_threshold')
self.randomseed = 1234
if cfg.has_option('GENERAL', 'seed'):
self.randomseed = cfg.getint('GENERAL', 'seed')
self.gamma = 0.99
if cfg.has_option('dqnpolicy', 'gamma'):
self.gamma = cfg.getfloat('dqnpolicy', 'gamma')
self.regularisation = 'l2'
if cfg.has_option('dqnpolicy', 'regularisation'):
self.regularisation = cfg.get('dqnpolicy', 'regularisation')
self.learning_rate = 0.001
if cfg.has_option('dqnpolicy', 'learning_rate'):
self.learning_rate = cfg.getfloat('dqnpolicy', 'learning_rate')
self.exploration_type = 'e-greedy' # Boltzman
if cfg.has_option('dqnpolicy', 'exploration_type'):
self.exploration_type = cfg.get('dqnpolicy', 'exploration_type')
self.episodeNum = 1000
if cfg.has_option('dqnpolicy', 'episodeNum'):
self.episodeNum = cfg.getfloat('dqnpolicy', 'episodeNum')
self.maxiter = 5000
if cfg.has_option('dqnpolicy', 'maxiter'):
self.maxiter = cfg.getfloat('dqnpolicy', 'maxiter')
self.curiosityreward = False
if cfg.has_option('eval', 'curiosityreward'):
self.curiosityreward = cfg.getboolean('eval', 'curiosityreward')
self.epsilon = 1
if cfg.has_option('dqnpolicy', 'epsilon'):
self.epsilon = cfg.getfloat('dqnpolicy', 'epsilon')
if not self.curiosityreward: # no eps-greedy exploration when curious expl. is used
self.epsilon_start = 1
if cfg.has_option('dqnpolicy', 'epsilon_start'):
self.epsilon_start = cfg.getfloat('dqnpolicy', 'epsilon_start')
else:
self.epsilon_start = 1
if cfg.has_option('dqnpolicy', 'epsilon_start'):
self.epsilon_start = cfg.getfloat('dqnpolicy', 'epsilon_start')
self.epsilon_end = 1
if cfg.has_option('dqnpolicy', 'epsilon_end'):
self.epsilon_end = cfg.getfloat('dqnpolicy', 'epsilon_end')
self.priorProbStart = 1.0
if cfg.has_option('dqnpolicy', 'prior_sample_prob_start'):
self.priorProbStart = cfg.getfloat('dqnpolicy',
'prior_sample_prob_start')
self.priorProbEnd = 0.1
if cfg.has_option('dqnpolicy', 'prior_sample_prob_end'):
self.priorProbEnd = cfg.getfloat('dqnpolicy',
'prior_sample_prob_end')
self.policyfeatures = []
if cfg.has_option('dqnpolicy', 'features'):
logger.info('Features: ' + str(cfg.get('dqnpolicy', 'features')))
self.policyfeatures = json.loads(cfg.get('dqnpolicy', 'features'))
self.max_k = 5
if cfg.has_option('dqnpolicy', 'max_k'):
self.max_k = cfg.getint('dqnpolicy', 'max_k')
self.learning_algorithm = 'drl'
if cfg.has_option('dqnpolicy', 'learning_algorithm'):
self.learning_algorithm = cfg.get('dqnpolicy',
'learning_algorithm')
logger.info('Learning algorithm: ' + self.learning_algorithm)
self.minibatch_size = 32
if cfg.has_option('dqnpolicy', 'minibatch_size'):
self.minibatch_size = cfg.getint('dqnpolicy', 'minibatch_size')
self.capacity = 1000
if cfg.has_option('dqnpolicy', 'capacity'):
self.capacity = cfg.getint('dqnpolicy', 'capacity')
self.replay_type = 'vanilla'
if cfg.has_option('dqnpolicy', 'replay_type'):
self.replay_type = cfg.get('dqnpolicy', 'replay_type')
self.architecture = 'vanilla'
if cfg.has_option('dqnpolicy', 'architecture'):
self.architecture = cfg.get('dqnpolicy', 'architecture')
self.q_update = 'single'
if cfg.has_option('dqnpolicy', 'q_update'):
self.q_update = cfg.get('dqnpolicy', 'q_update')
self.h1_size = 130
if cfg.has_option('dqnpolicy', 'h1_size'):
self.h1_size = cfg.getint('dqnpolicy', 'h1_size')
self.h2_size = 50
if cfg.has_option('dqnpolicy', 'h2_size'):
self.h2_size = cfg.getint('dqnpolicy', 'h2_size')
self.save_step = 200
if cfg.has_option('policy', 'save_step'):
self.save_step = cfg.getint('policy', 'save_step')
self.importance_sampling = 'soft'
if cfg.has_option('dqnpolicy', 'importance_sampling'):
self.importance_sampling = cfg.get('dqnpolicy',
'importance_sampling')
self.train_iters_per_episode = 1
if cfg.has_option('dqnpolicy', 'train_iters_per_episode'):
self.train_iters_per_episode = cfg.getint(
'dqnpolicy', 'train_iters_per_episode')
self.training_frequency = 2
if cfg.has_option('dqnpolicy', 'training_frequency'):
self.training_frequency = cfg.getint('dqnpolicy',
'training_frequency')
# domain specific parameter settings (overrides general policy parameter settings)
if cfg.has_option('dqnpolicy_' + domainString, 'n_in'):
self.n_in = cfg.getint('dqnpolicy_' + domainString, 'n_in')
if cfg.has_option('dqnpolicy_' + domainString, 'actor_lr'):
self.actor_lr = cfg.getfloat('dqnpolicy_' + domainString,
'actor_lr')
if cfg.has_option('dqnpolicy_' + domainString, 'critic_lr'):
self.critic_lr = cfg.getfloat('dqnpolicy_' + domainString,
'critic_lr')
if cfg.has_option('dqnpolicy_' + domainString, 'delta'):
self.delta = cfg.getfloat('dqnpolicy_' + domainString, 'delta')
if cfg.has_option('dqnpolicy_' + domainString, 'beta'):
self.alpha = cfg.getfloat('dqnpolicy_' + domainString, 'beta')
if cfg.has_option('dqnpolicy_' + domainString, 'is_threshold'):
self.c = cfg.getfloat('dqnpolicy_' + domainString, 'is_threshold')
if cfg.has_option('dqnpolicy_' + domainString, 'gamma'):
self.gamma = cfg.getfloat('dqnpolicy_' + domainString, 'gamma')
if cfg.has_option('dqnpolicy_' + domainString, 'regularisation'):
self.regularisation = cfg.get('dqnpolicy_' + domainString,
'regulariser')
if cfg.has_option('dqnpolicy_' + domainString, 'learning_rate'):
self.learning_rate = cfg.getfloat('dqnpolicy_' + domainString,
'learning_rate')
if cfg.has_option('dqnpolicy_' + domainString, 'exploration_type'):
self.exploration_type = cfg.get('dqnpolicy_' + domainString,
'exploration_type')
if cfg.has_option('dqnpolicy_' + domainString, 'episodeNum'):
self.episodeNum = cfg.getfloat('dqnpolicy_' + domainString,
'episodeNum')
if cfg.has_option('dqnpolicy_' + domainString, 'maxiter'):
self.maxiter = cfg.getfloat('dqnpolicy_' + domainString, 'maxiter')
if cfg.has_option('dqnpolicy_' + domainString, 'epsilon'):
self.epsilon = cfg.getfloat('dqnpolicy_' + domainString, 'epsilon')
if cfg.has_option('dqnpolicy_' + domainString, 'epsilon_start'):
self.epsilon_start = cfg.getfloat('dqnpolicy_' + domainString,
'epsilon_start')
if cfg.has_option('dqnpolicy_' + domainString, 'epsilon_end'):
self.epsilon_end = cfg.getfloat('dqnpolicy_' + domainString,
'epsilon_end')
if cfg.has_option('dqnpolicy_' + domainString,
'prior_sample_prob_start'):
self.priorProbStart = cfg.getfloat('dqnpolicy_' + domainString,
'prior_sample_prob_start')
if cfg.has_option('dqnpolicy_' + domainString,
'prior_sample_prob_end'):
self.priorProbEnd = cfg.getfloat('dqnpolicy_' + domainString,
'prior_sample_prob_end')
if cfg.has_option('dqnpolicy_' + domainString, 'features'):
logger.info('Features: ' +
str(cfg.get('dqnpolicy_' + domainString, 'features')))
self.policyfeatures = json.loads(
cfg.get('dqnpolicy_' + domainString, 'features'))
if cfg.has_option('dqnpolicy_' + domainString, 'max_k'):
self.max_k = cfg.getint('dqnpolicy_' + domainString, 'max_k')
if cfg.has_option('dqnpolicy_' + domainString, 'learning_algorithm'):
self.learning_algorithm = cfg.get('dqnpolicy_' + domainString,
'learning_algorithm')
logger.info('Learning algorithm: ' + self.learning_algorithm)
if cfg.has_option('dqnpolicy_' + domainString, 'minibatch_size'):
self.minibatch_size = cfg.getint('dqnpolicy_' + domainString,
'minibatch_size')
if cfg.has_option('dqnpolicy_' + domainString, 'capacity'):
self.capacity = cfg.getint('dqnpolicy_' + domainString, 'capacity')
if cfg.has_option('dqnpolicy_' + domainString, 'replay_type'):
self.replay_type = cfg.get('dqnpolicy_' + domainString,
'replay_type')
if cfg.has_option('dqnpolicy_' + domainString, 'architecture'):
self.architecture = cfg.get('dqnpolicy_' + domainString,
'architecture')
if cfg.has_option('dqnpolicy_' + domainString, 'q_update'):
self.q_update = cfg.get('dqnpolicy_' + domainString, 'q_update')
if cfg.has_option('dqnpolicy_' + domainString, 'h1_size'):
self.h1_size = cfg.getint('dqnpolicy_' + domainString, 'h1_size')
if cfg.has_option('dqnpolicy_' + domainString, 'h2_size'):
self.h2_size = cfg.getint('dqnpolicy_' + domainString, 'h2_size')
if cfg.has_option('policy_' + domainString, 'save_step'):
self.save_step = cfg.getint('policy_' + domainString, 'save_step')
if cfg.has_option('dqnpolicy_' + domainString, 'importance_sampling'):
self.importance_sampling = cfg.get('dqnpolicy_' + domainString,
'importance_sampling')
if cfg.has_option('dqnpolicy_' + domainString,
'train_iters_per_episode'):
self.train_iters_per_episode = cfg.getint(
'dqnpolicy_' + domainString, 'train_iters_per_episode')
if cfg.has_option('dqnpolicy_' + domainString, 'training_frequency'):
self.training_frequency = cfg.getint('dqnpolicy_' + domainString,
'training_frequency')
self.episode_ct = 0
self.episode_ave_max_q = []
self.mu_prob = 0. # behavioral policy
os.environ["CUDA_VISIBLE_DEVICES"] = ""
# init session
self.sess = tf.Session()
with tf.device("/cpu:0"):
np.random.seed(self.randomseed)
tf.set_random_seed(self.randomseed)
# initialise an replay buffer
if self.replay_type == 'vanilla':
self.episodes[self.domainString] = ReplayBufferEpisode(
self.capacity, self.minibatch_size, self.randomseed)
elif self.replay_type == 'prioritized':
self.episodes[self.domainString] = ReplayPrioritisedEpisode(
self.capacity, self.minibatch_size, self.randomseed)
#replay_buffer = ReplayBuffer(self.capacity, self.randomseed)
#self.episodes = []
self.samplecount = 0
self.episodecount = 0
# construct the models
self.state_dim = self.n_in
self.summaryaction = SummaryAction.SummaryAction(domainString)
self.action_dim = len(self.summaryaction.action_names)
action_bound = len(self.summaryaction.action_names)
self.stats = [0 for _ in range(self.action_dim)]
self.global_mu = [0. for _ in range(self.action_dim)]
self.acer = acer.ACERNetwork(self.sess, self.state_dim, self.action_dim, \
self.critic_lr, self.delta, self.c, self.alpha, self.h1_size, self.h2_size, self.is_training, self.randomseed)
# when all models are defined, init all variables
init_op = tf.global_variables_initializer()
self.sess.run(init_op)
self.loadPolicy(self.in_policy_file)
print 'loaded replay size: ', self.episodes[
self.domainString].size()
#improvement==================================
#initial
if self.intrinsic_reward_method == 'vime':
self.vime_model = vime(self.state_dim, self.action_dim)
self.vime_model.load_model('model/vime_model/' +
self.in_policy_file)
elif self.intrinsic_reward_method == 'cme':
self.cme_model = cme(self.state_dim, self.action_dim)
self.cme_model.load_model('model/cme_model/' +
self.in_policy_file)
elif self.intrinsic_reward_method == 'scme':
self.scme_model = scme(self.state_dim, self.action_dim)
self.scme_model.load_model('model/scme_model/' +
self.in_policy_file)
