def __init__(self, observation_space, action_space, config, dqn_config):
BaselinePolicy.__init__(self, observation_space, action_space, config)
self.dqn_config = dqn_config
self.epsilon = 1
mixed_dqn_config = mixed_dqn_net_config_example.copy()
mixed_dqn_config.update({
"controllable_state_num" : int(np.product(observation_space.shape)),
"action_num": int(action_space.n),
"uncontrollable_state_num": 31,
"uncontrollable_pred_num": 3,
})
self.num_states = int(np.product(observation_space.shape))
self.num_actions = int(action_space.n)
print(f'dqn state space:{self.num_states}, action space:{self.num_actions}')
self.use_unc_part = dqn_config['use_unc_part']
self.eval_net = mixed_dqn_net(mixed_dqn_net_config_example)
self.target_net = mixed_dqn_net(mixed_dqn_net_config_example)
self.target_net.load_state_dict(self.eval_net.state_dict())
self.learn_step_counter = 0
self.memory = replay_memory(dqn_config['replay_capacity'])
self.optimizer = torch.optim.Adam(self.eval_net.parameters(), lr=dqn_config['lr'])
self.loss_func = nn.SmoothL1Loss()
self.rand_action = 0
self.greedy_action = 0
def __init__(self, observation_space, action_space, config, dqn_config):
BaselinePolicy.__init__(self, observation_space, action_space, config)
self.dqn_config = dqn_config
self.epsilon = 1
self.num_states = int(np.product(observation_space.shape))
self.num_actions = int(action_space.n)
print(
f'dqn state space:{self.num_states}, action space:{self.num_actions}'
)
self.pred_head = dqn_config['pred']
self.eval_net = dqn_net(self.num_states, self.num_actions,
self.pred_head)
self.target_net = dqn_net(self.num_states, self.num_actions,
self.pred_head)
self.target_net.load_state_dict(self.eval_net.state_dict())
self.learn_step_counter = 0
self.memory = replay_memory(dqn_config['replay_capacity'])
self.optimizer = torch.optim.Adam(self.eval_net.parameters(),
lr=dqn_config['lr'])
self.loss_func = nn.SmoothL1Loss()
self.rand_action = 0
self.greedy_action = 0
def __init__(self, observation_space, action_space, config, dqn_config):
BaselinePolicy.__init__(self, observation_space, action_space, config)
self.dqn_config = dqn_config
self.epsilon = 1
self.Vmin = -1
self.Vmax = 1
self.atoms = 51
self.device = torch.device('cpu')
self.num_states = int(np.product(observation_space.shape))
self.num_actions = int(action_space.n)
print(
f'dqn state space:{self.num_states}, action space:{self.num_actions}'
)
self.eval_net = c51_net(self.num_states, self.num_actions, self.atoms)
self.target_net = c51_net(self.num_states, self.num_actions,
self.atoms)
self.target_net.load_state_dict(self.eval_net.state_dict())
self.learn_step_counter = 0
self.memory = replay_memory(dqn_config['replay_capacity'])
self.optimizer = torch.optim.Adam(self.eval_net.parameters(),
lr=dqn_config['lr'])
self.support = torch.linspace(self.Vmin, self.Vmax, self.atoms)
def __init__(self, observation_space, action_space, config, dqn_config):
BaselinePolicy.__init__(self, observation_space, action_space, config)
self.dqn_config = dqn_config
self.epsilon = 1
mixed_dqn_config = mixed_dqn_net_config_example.copy()
mixed_dqn_config.update({
"controllable_state_num":
int(np.product(observation_space.shape)),
"action_num":
int(action_space.n),
"uncontrollable_state_num":
21 * 7, # 31
"uncontrollable_pred_num":
6,
'fixed_uncontrollable_param':
dqn_config['fixed_uncontrollable_param'],
'uncontrollable_use_cnn':
dqn_config['use_cnn_state'],
'embeddingmerge':
dqn_config['embeddingmerge'],
'activation_func':
dqn_config['activation_func'],
'use_bn':
dqn_config['use_bn'],
})
self.num_states = int(np.product(observation_space.shape))
self.num_actions = int(action_space.n)
print(
f'dqn state space:{self.num_states}, action space:{self.num_actions}'
)
self.use_unc_part = dqn_config['use_unc_part']
#self.pre_train = dqn_config['pretrain']
self.fixed_uncontrollable_param = dqn_config['pretrain']
if dqn_config['use_cnn_state']:
dqn_net = mixed_dqn_unc_cnn_net
else:
dqn_net = mixed_dqn_net
self.eval_net = dqn_net(mixed_dqn_config)
self.target_net = dqn_net(mixed_dqn_config)
self.target_net.load_state_dict(self.eval_net.state_dict())
self.learn_step_counter = 0
self.memory = replay_memory(dqn_config['replay_capacity'])
self.eval_memory = replay_memory(dqn_config['replay_capacity'])
self.optimizer = torch.optim.Adam(
self.eval_net.parameters(),
lr=dqn_config['lr'],
weight_decay=dqn_config['weight_decay'])
self.loss_func = nn.SmoothL1Loss()
self.rand_action = 0
self.greedy_action = 0
self.evaluation = False # only for epsilon-greedy
# augmentation setting
self.train_augmentation = dqn_config['train_augmentation']
self.demand_augmentation = demand_augmentation(
noise_type=dqn_config['train_augmentation'],
noise_scale=dqn_config['noise_scale'],
sparse_scale=dqn_config['sparse_scale'])
def __init__(self, observation_space, action_space, config):
BaselinePolicy.__init__(self, observation_space, action_space, config)