def __init__(
self,
env,
qf,
vf,
sac_kwargs,
tdm_kwargs,
base_kwargs,
policy=None,
replay_buffer=None,
give_terminal_reward=False,
):
SoftActorCritic.__init__(self,
env=env,
policy=policy,
qf=qf,
vf=vf,
replay_buffer=replay_buffer,
**sac_kwargs,
**base_kwargs)
TemporalDifferenceModel.__init__(self, **tdm_kwargs)
action_space_diff = (self.env.action_space.high -
self.env.action_space.low)
# TODO(vitchyr): Maybe add this to the main SAC code.
terminal_reward = 0
for dim in range(action_space_diff.size):
terminal_reward += (-np.log(1. / action_space_diff[dim]))
self.terminal_bonus = float(terminal_reward)
self.give_terminal_reward = give_terminal_reward
def __init__(self,
*args,
observation_key=None,
desired_goal_key=None,
**kwargs):
HER.__init__(
self,
observation_key=observation_key,
desired_goal_key=desired_goal_key,
)
SoftActorCritic.__init__(self, *args, **kwargs)
assert isinstance(self.replay_buffer, ObsDictRelabelingBuffer)