def __init__(self):
super().__init__()
self.input_embedders_parameters = {'observation': InputEmbedderParameters()}
self.middleware_parameters = FCMiddlewareParameters()
self.heads_parameters = [VHeadParameters(loss_weight=0.5), PolicyHeadParameters(loss_weight=1.0)]
self.optimizer_type = 'Adam'
self.clip_gradients = 40.0
self.async_training = True
def __init__(self):
super().__init__()
self.input_embedders_parameters = {
'observation': InputEmbedderParameters(activation_function='tanh')
}
self.middleware_parameters = FCMiddlewareParameters(
activation_function='tanh')
self.heads_parameters = [VHeadParameters()]
self.async_training = True
self.l2_regularization = 0
self.create_target_network = True
self.batch_size = 128
def __init__(self):
super().__init__()
self.input_embedders_parameters = {'observation': InputEmbedderParameters(activation_function='relu')}
self.middleware_parameters = FCMiddlewareParameters(activation_function='relu')
self.heads_parameters = [VHeadParameters(initializer='xavier')]
self.rescale_gradient_from_head_by_factor = [1]
self.optimizer_type = 'Adam'
self.batch_size = 256
self.async_training = False
self.learning_rate = 0.0003 # 3e-4 see appendix D in the paper
# tau is set in SoftActorCriticAlgorithmParameters.rate_for_copying_weights_to_target
self.create_target_network = True
def __init__(self):
super().__init__()
self.input_embedders_parameters = {
'observation': InputEmbedderParameters(batchnorm=True),
'action': InputEmbedderParameters(scheme=EmbedderScheme.Shallow)
}
self.middleware_parameters = FCMiddlewareParameters()
self.heads_parameters = [VHeadParameters()]
self.optimizer_type = 'Adam'
self.batch_size = 64
self.async_training = False
self.learning_rate = 0.001
self.create_target_network = True
self.shared_optimizer = True
self.scale_down_gradients_by_number_of_workers_for_sync_training = False
def __init__(self):
super().__init__()
self.input_embedders_parameters = {'observation': InputEmbedderParameters(activation_function='tanh')}
self.middleware_parameters = FCMiddlewareParameters(activation_function='tanh')
self.heads_parameters = [VHeadParameters(), PPOHeadParameters()]
self.batch_size = 64
self.optimizer_type = 'Adam'
self.clip_gradients = None
self.use_separate_networks_per_head = True
self.async_training = False
self.l2_regularization = 0
# The target network is used in order to freeze the old policy, while making updates to the new one
# in train_network()
self.create_target_network = True
self.shared_optimizer = True
self.scale_down_gradients_by_number_of_workers_for_sync_training = True