def action_space(self):
space = self._env.action_space
if not self._should_normalize_action:
return space
if isinstance(space, tools.ActTuple):
return tools.ActTuple((
spaces.Discrete(3), # offensive decision
# ball theta
spaces.Box(-np.ones(space[1].shape),
np.ones(space[1].shape),
dtype=np.float32),
# offense player DASH(power, direction)
spaces.Box(-np.ones(space[2].shape),
np.ones(space[2].shape),
dtype=np.float32),
# defense player DASH(power, direction)
spaces.Box(-np.ones(space[3].shape),
np.ones(space[3].shape),
dtype=np.float32)))
elif isinstance(space, tools.NDefActTuple):
return tools.NDefActTuple((
spaces.Discrete(3), # offensive decision
# ball theta
spaces.Box(-np.ones(space[1].shape),
np.ones(space[1].shape),
dtype=np.float32),
# offense player DASH(power, direction)
spaces.Box(-np.ones(space[2].shape),
np.ones(space[2].shape),
dtype=np.float32)))
def _set_action_space(self):
"""
Return
------
ActTuple(Discrete(3), Box(2), Box(5, 2), Box(5, 2))
"""
return tools.ActTuple((
spaces.Discrete(3), # offensive decision
# ball dir
spaces.Box(low=-1, high=1, shape=(2, ), dtype=np.float32),
# offense player DASH(x, y)
spaces.Box(low=-self.pl_max_power,
high=self.pl_max_power,
shape=(5, 2),
dtype=np.float32),
# defense player DASH(x, y)
spaces.Box(low=-self.pl_max_power,
high=self.pl_max_power,
shape=(5, 2),
dtype=np.float32)))
def _set_action_space(self):
"""
Return
------
ActTuple(Discrete(3), Box(), Box(5, 2), Box(5, 2))
"""
return tools.ActTuple((
spaces.Discrete(3), # offensive decision
# ball theta
spaces.Box(low=-np.pi, high=np.pi, shape=(), dtype=np.float32),
# offense player DASH(power, direction)
spaces.Box(low=np.array([[0, -np.pi] for _ in range(5)],
dtype=np.float32),
high=np.array([[self.pl_max_power, np.pi]
for _ in range(5)],
dtype=np.float32),
dtype=np.float32),
# defense player DASH(power, direction)
spaces.Box(low=np.array([[0, -np.pi] for _ in range(5)],
dtype=np.float32),
high=np.array([[self.pl_max_power, np.pi]
for _ in range(5)],
dtype=np.float32),
dtype=np.float32)))