def worker_reward_fn(states, goals, next_states):
return negative_distance(states=states,
state_indices=self.goal_indices,
goals=goals,
next_states=next_states,
relative_context=relative_goals,
offset=0.0)
def intrinsic_reward_fn(states, goals, next_states):
return negative_distance(
states=states[self.goal_indices] / scale,
goals=goals / scale,
next_states=next_states[self.goal_indices] / scale,
relative_context=relative_goals,
offset=0.0) + offset
def contextual_reward(states, goals, next_states):
return negative_distance(states=states,
goals=goals,
next_states=next_states,
state_indices=[0, 1, 2],
relative_context=False,
offset=0.0,
reward_scales=REWARD_SCALE)
def contextual_reward(states, goals, next_states):
return negative_distance(states=states,
goals=goals,
next_states=next_states,
state_indices=[0, 1],
relative_context=False,
offset=0.0,
reward_scales=1 / 7.2,
output_activation=np.exp)
def test_negative_distance(self):
a = np.array([1, 2, 10])
b = np.array([1, 2])
c = negative_distance(b, b, a, goal_indices=[1, 2])
self.assertEqual(c, -8.062257748304752)
