def __init__(self):
super().__init__()
self.input_embedders_parameters = {'observation': InputEmbedderParameters()}
self.middleware_parameters = FCMiddlewareParameters()
self.heads_parameters = [PolicyHeadParameters()]
self.loss_weights = [1.0]
self.async_training = True
def __init__(self):
super().__init__()
self.input_embedders_parameters = {'observation': InputEmbedderParameters()}
self.middleware_parameters = FCMiddlewareParameters(scheme=MiddlewareScheme.Medium)
self.heads_parameters = [PolicyHeadParameters()]
self.optimizer_type = 'Adam'
self.batch_size = 32
self.replace_mse_with_huber_loss = False
self.create_target_network = False
def __init__(self):
super().__init__()
self.input_embedders_parameters = {
'observation': InputEmbedderParameters()
}
self.middleware_parameters = FCMiddlewareParameters()
self.heads_parameters = [VHeadParameters(), PolicyHeadParameters()]
self.loss_weights = [0.5, 1.0]
self.rescale_gradient_from_head_by_factor = [1, 1]
self.optimizer_type = 'Adam'
self.clip_gradients = 40.0
self.async_training = True
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()
}
self.middleware_parameters = FCMiddlewareParameters()
self.heads_parameters = [VHeadParameters(), PolicyHeadParameters()]
self.loss_weights = [0.5, 1.0]
self.sil_loss_weights = [0.5 * 0.01,
1.0] # called beta^SIL in the paper
self.rescale_gradient_from_head_by_factor = [1, 1]
self.optimizer_type = 'Adam'
self.clip_gradients = 40.0
self.batch_size = 32 # = 512 / 16 workers (since training is synchronous)
self.async_training = False # A2C
self.shared_optimizer = True