def get_sc2_action(self, gym_action) -> List[FunctionCall]:
if FUNCTIONS.Move_screen.id not in self.available_actions:
return [FUNCTIONS.select_army("select")]
# For restricting the action space for A2C, does not break DQN
gym_action %= self.action_space.n
# Find mean coordinates of all currently active player units
player_units_xy = [(unit.x, unit.y)
for unit in self.state["player_units"]]
arr = np.asarray(player_units_xy)
length = arr.shape[0]
x_sum = np.sum(arr[:, 0])
y_sum = np.sum(arr[:, 1])
centroid = (int(x_sum / length), int(y_sum / length))
# 0: Up
# 1: Down
# 2: Left
# 3: Right
# 4: Up + Left
# 5: Up + Right
# 6: Down + Left
# 7: Down + Right
target_xy = list(centroid)
x_max = self.screen_shape[0] - 1
y_max = self.screen_shape[1] - 1
# Determine target position, diff => min(abs(x_diff), abs(y_diff))
if gym_action == 0: target_xy[1] = 0
elif gym_action == 1: target_xy[1] = y_max
elif gym_action == 2: target_xy[0] = 0
elif gym_action == 3: target_xy[0] = x_max
elif gym_action == 4:
diff = min(centroid[0], centroid[1])
target_xy = [target_xy[0] - diff, target_xy[1] - diff]
elif gym_action == 5:
diff = min(x_max - centroid[0], centroid[1])
target_xy = [target_xy[0] + diff, target_xy[1] - diff]
elif gym_action == 6:
diff = min(centroid[0], y_max - centroid[1])
target_xy = [target_xy[0] - diff, target_xy[1] + diff]
elif gym_action == 7:
diff = min(x_max - centroid[0], y_max - centroid[1])
target_xy = [target_xy[0] + diff, target_xy[1] + diff]
# Assign action function
# Move_screen
action = FUNCTIONS.Move_screen("now", target_xy)
return [action]
def get_sc2_action(self, gym_action) -> List[FunctionCall]:
# Get coords by unravelling action.
# How unravel works:
# Ref: https://www.quora.com/What-is-a-simple-intuitive-example-for-the-unravel_index-in-Python
coords = np.unravel_index(gym_action, (self.screen_shape[0], self.screen_shape[1]))
# PySC2 uses different conventions for observations (y,x) and actions (x,y)
action = FUNCTIONS.Move_screen("now", coords[::-1]) # ::-1 reverses the tuple i.e. (1,2) becomes (2,1)
if action.function not in self.available_actions:
# logger.warning("Attempted unavailable action {}.".format(action))
action = FUNCTIONS.select_army("select")
return [action]
def get_sc2_action(self, gym_action) -> List[FunctionCall]:
if FUNCTIONS.Move_screen.id not in self.available_actions:
return [FUNCTIONS.select_army("select")]
# 0 = no-op
if gym_action == 0:
return [FUNCTIONS.no_op()]
player_unit_xy = [
self.state["player_units"][0].x, self.state["player_units"][0].y
]
target_xy = player_unit_xy
# 0: No-op
# 1: Up
# 2: Down
# 3: Left
# 4: Right
# 5: Up + Left
# 6: Up + Right
# 7: Down + Left
# 8: Down + Right
# Determine target position
if gym_action in (1, 5, 6):
# Up
target_xy[1] = max(0, player_unit_xy[1] - 1)
if gym_action in (2, 7, 8):
# Down
target_xy[1] = min(self.screen_shape[1] - 1, player_unit_xy[1] + 1)
if gym_action in (3, 5, 7):
# Left
target_xy[0] = max(0, player_unit_xy[0] - 1)
if gym_action in (4, 6, 8):
# Right
target_xy[0] = min(self.screen_shape[0] - 1, player_unit_xy[0] + 1)
# Assign action function
# Move_screen
action = FUNCTIONS.Move_screen("now", target_xy)
return [action]
def step(self, obs, *args):
if obs.last():
self.reset()
marine = get_units_by_type(obs.observation, units.Terran.Marine)[0]
beacons = get_units_by_type(obs.observation, 317)
targets = set((b[1]['x'], b[1]['y']) for b in beacons)
if not targets:
return FUNCTIONS.no_op()
if self.target not in targets:
self.target = None
if not marine[1]['is_selected']:
return select_unit(marine)
if self.target:
return FUNCTIONS.no_op()
distance = get_distances(marine, beacons)
beacon = beacons[np.argmin(distance)]
self.target = (beacon[1]['x'], beacon[1]['y'])
return FUNCTIONS.Move_screen("now", (beacon[1].x, beacon[1].y))
def moveScreen(x1, y1):
action = Action()
action.addCommand(FUNCTIONS.Move_screen("now", [x1, y1]))
return action
def _move_marine_to_minarel(self, marine, mineral):
# todo , move to closet position randomly
spatial_action = FUNCTIONS.Move_screen("now",
(mineral[1].x, mineral[1].y))
return spatial_action