def test_discretized(self):
n = 11
min_action = -10.0
max_action = 10.0
space = Discretized(n, min_action, max_action)
random_action = space.sample()
self.assertGreaterEqual(random_action, 0)
self.assertLess(random_action, n)
expected_value = min_action
step = (max_action - min_action) / (n - 1)
for action in range(n):
continuous_action = space.to_continuous(action)
self.assertAlmostEqual(continuous_action, expected_value)
expected_value += step
def doom_action_space_full_discretized(with_use=False):
"""
MOVE_FORWARD
MOVE_BACKWARD
MOVE_RIGHT
MOVE_LEFT
SELECT_WEAPON1
SELECT_WEAPON2
SELECT_WEAPON3
SELECT_WEAPON4
SELECT_WEAPON5
SELECT_WEAPON6
SELECT_WEAPON7
ATTACK
SPEED
TURN_LEFT_RIGHT_DELTA
"""
spaces = [
Discrete(3), # noop, forward, backward
Discrete(3), # noop, move right, move left
Discrete(8), # noop, select weapons 1-7
Discrete(2), # noop, attack
Discrete(2), # noop, sprint
]
if with_use:
spaces.append(Discrete(2)) # noop, use
spaces.append(Discretized(
21, min_action=-12.5,
max_action=12.5)) # turning using discretized continuous control
return gym.spaces.Tuple(spaces)
def doom_action_space_discretized_no_weap():
return gym.spaces.Tuple((
Discrete(3), # noop, forward, backward
Discrete(3), # noop, move right, move left
Discrete(2), # noop, attack
Discrete(2), # noop, sprint
Discretized(
11, min_action=-10.0,
max_action=10.0), # turning using discretized continuous control
))