def worker(remote, map_name, i):
with sc2_env.SC2Env(map_name=map_name,
step_mul=1,
screen_size_px=(32, 32),
minimap_size_px=(32, 32)) as env:
available_actions = None
result = None
group_list = []
while True:
cmd, data = remote.recv()
if cmd == 'step':
# if(common.check_group_list(env, result)):
# result, xy_per_marine = common.init(env,result)
reward = 0
if (len(group_list) == 0
or common.check_group_list(env, result)):
print("init group list")
result, xy_per_marine = common.init(env, result)
group_list = common.update_group_list(result)
action1 = data[0][0]
action2 = data[0][1]
func = actions.FUNCTIONS[action1[0]]
#print("agent(",i," ) action : ", action1, " func : ", func)
func = actions.FUNCTIONS[action2[0]]
#print("agent(",i," ) action : ", action2, " func : ", func, "xy :", action2[1][1])
x, y = action2[1][1]
move = True
if (x == 0 and y == 0):
move = False
result = env.step(actions=[action1])
reward += result[0].reward
done = result[0].step_type == environment.StepType.LAST
if (331 in available_actions and move and not done):
try:
result = env.step(actions=[action2])
reward += result[0].reward
done = result[0].step_type == environment.StepType.LAST
except Exception as e:
print("e :", e)
ob = (result[0].observation["screen"]
[_PLAYER_RELATIVE:_PLAYER_RELATIVE + 1] == 3).astype(
int) # (1, 32, 32)
selected = result[0].observation["screen"][
_SELECTED:_SELECTED + 1] # (1, 32, 32)
# extra = np.zeros((1, 32, 32))
control_groups = result[0].observation["control_groups"]
army_count = env._obs[0].observation.player_common.army_count
# extra[0,0,0] = army_count
# for id, group in enumerate(control_groups):
# control_group_id = id
# unit_id = group[0]
# count = group[1]
# #print("control_group_id :", control_group_id, " unit_id :", unit_id, " count :", count)
# extra[0,1, control_group_id] = unit_id
# extra[0,2, control_group_id] = count
#ob = np.append(ob, selected, axis=0) # (2, 32, 32)
#ob = np.append(ob, extra, axis=0) # (3, 32, 32)
available_actions = result[0].observation["available_actions"]
info = result[0].observation["available_actions"]
if done:
result = env.reset()
if (len(group_list) == 0
or common.check_group_list(env, result)):
print("init group list")
result, xy_per_marine = common.init(env, result)
group_list = common.update_group_list(result)
# ob = result[0].observation["screen"]
# reward = result[0].reward
# done = result[0].step_type == environment.StepType.LAST
info = result[0].observation["available_actions"]
remote.send((ob, reward, done, info, army_count,
control_groups, selected, xy_per_marine))
elif cmd == 'reset':
result = env.reset()
reward = 0
if (len(group_list) == 0
or common.check_group_list(env, result)):
print("init group list")
result, xy_per_marine = common.init(env, result)
group_list = common.update_group_list(result)
reward += result[0].reward
ob = (result[0].observation["screen"]
[_PLAYER_RELATIVE:_PLAYER_RELATIVE + 1] == 3).astype(int)
selected = result[0].observation["screen"][
_SELECTED:_SELECTED + 1] # (1, 32, 32)
# extra = np.zeros((1, 32, 32))
control_groups = result[0].observation["control_groups"]
army_count = env._obs[0].observation.player_common.army_count
# extra[0,0,0] = army_count
# for id, group in enumerate(control_groups):
# control_group_id = id
# unit_id = group[0]
# count = group[1]
# #print("control_group_id :", control_group_id, " unit_id :", unit_id, " count :", count)
# extra[0,1, control_group_id] = unit_id
# extra[0,2, control_group_id] = count
# ob = np.append(ob, selected, axis=0) # (2, 32, 32)
# ob = np.append(ob, extra, axis=0) # (3, 32, 32)
done = result[0].step_type == environment.StepType.LAST
info = result[0].observation["available_actions"]
available_actions = result[0].observation["available_actions"]
remote.send((ob, reward, done, info, army_count,
control_groups, selected, xy_per_marine))
elif cmd == 'close':
remote.close()
break
elif cmd == 'get_spaces':
remote.send((env.action_spec().functions[data], ""))
elif cmd == "action_spec":
remote.send((env.action_spec().functions[data]))
else:
raise NotImplementedError
def step(self, obs):
super(CollectMineralShards, self).step(obs)
if (len(self.group_list) == 0
or common.check_group_list(self.env, [obs])):
print("init group list")
obs = common.init(self.env, [obs])[0]
self.group_list = common.update_group_list([obs])
#print("group_list ", self.group_list)
player_relative = obs.observation["screen"][_PLAYER_RELATIVE]
neutral_y, neutral_x = (player_relative == _PLAYER_NEUTRAL).nonzero()
player_y, player_x = (player_relative == _PLAYER_FRIENDLY).nonzero()
if not neutral_y.any() or not player_y.any():
return actions.FunctionCall(_NO_OP, [])
r = random.randint(0, 1)
if _MOVE_SCREEN in obs.observation["available_actions"] and r == 0:
selected = obs.observation["screen"][_SELECTED]
player_y, player_x = (selected == _PLAYER_FRIENDLY).nonzero()
player = [int(player_x.mean()), int(player_y.mean())]
points = [player]
closest, min_dist = None, None
other_dest = None
my_dest = None
if ("0" in self.dest_per_marine and "1" in self.dest_per_marine):
if (self.group_id == 0):
my_dest = self.dest_per_marine["0"]
other_dest = self.dest_per_marine["1"]
elif (self.group_id == 1):
other_dest = self.dest_per_marine["0"]
my_dest = self.dest_per_marine["1"]
for p in zip(neutral_x, neutral_y):
if (other_dest):
dist = numpy.linalg.norm(
numpy.array(other_dest) - numpy.array(p))
if (dist < 5):
#print("continue since partner will take care of it ", p)
continue
pp = [p[0] // 2 * 2, p[1] // 2 * 2]
if (pp not in points):
points.append(pp)
dist = numpy.linalg.norm(numpy.array(player) - numpy.array(p))
if not min_dist or dist < min_dist:
closest, min_dist = p, dist
solve_tsp = False
if (my_dest):
dist = numpy.linalg.norm(
numpy.array(player) - numpy.array(my_dest))
if (dist < 2):
solve_tsp = True
if (my_dest is None):
solve_tsp = True
if (len(points) < 2):
solve_tsp = False
if (solve_tsp):
# function for printing best found solution when it is found
from time import clock
init = clock()
def report_sol(obj, s=""):
print("cpu:%g\tobj:%g\ttour:%s" % \
(clock(), obj, s))
#print("points: %s" % points)
n, D = mk_matrix(points, distL2)
# multi-start local search
#print("random start local search:", n)
niter = 50
tour, z = multistart_localsearch(niter, n, D)
#print("best found solution (%d iterations): z = %g" % (niter, z))
#print(tour)
left, right = None, None
for idx in tour:
if (tour[idx] == 0):
if (idx == len(tour) - 1):
#print("optimal next : ", tour[0])
right = points[tour[0]]
left = points[tour[idx - 1]]
elif (idx == 0):
#print("optimal next : ", tour[idx+1])
right = points[tour[idx + 1]]
left = points[tour[len(tour) - 1]]
else:
#print("optimal next : ", tour[idx+1])
right = points[tour[idx + 1]]
left = points[tour[idx - 1]]
left_d = numpy.linalg.norm(
numpy.array(player) - numpy.array(left))
right_d = numpy.linalg.norm(
numpy.array(player) - numpy.array(right))
if (right_d > left_d):
closest = left
else:
closest = right
#print("optimal next :" , closest)
self.dest_per_marine[str(self.group_id)] = closest
#print("dest_per_marine", self.dest_per_marine)
#dest_per_marine {'0': [56, 26], '1': [52, 6]}
if (closest is None):
return actions.FunctionCall(_NO_OP, [])
return actions.FunctionCall(_MOVE_SCREEN, [_NOT_QUEUED, closest])
elif (len(self.group_list) > 0):
player_p = []
for p in zip(player_x, player_y):
if (p not in player_p):
player_p.append(p)
self.group_id = random.randint(0, len(self.group_list) - 1)
return actions.FunctionCall(
_SELECT_CONTROL_GROUP,
[[_CONTROL_GROUP_RECALL], [int(self.group_id)]])
else:
return actions.FunctionCall(_NO_OP, [])
def step(self, obs):
super(CollectMineralShards, self).step(obs)
if (len(self.group_list) == 0
or common.check_group_list(self.env, [obs])):
print("init group list")
obs, xy_per_marine = common.init(self.env, [obs])
obs = obs[0]
self.group_list = common.update_group_list([obs])
#print("group_list ", self.group_list)
player_relative = obs.observation["screen"][_PLAYER_RELATIVE]
neutral_y, neutral_x = (player_relative == _PLAYER_NEUTRAL).nonzero()
player_y, player_x = (player_relative == _PLAYER_FRIENDLY).nonzero()
if not neutral_y.any() or not player_y.any():
return actions.FunctionCall(_NO_OP, [])
r = random.randint(0, 1)
if _MOVE_SCREEN in obs.observation["available_actions"] and r == 0:
selected = obs.observation["screen"][_SELECTED]
player_y, player_x = (selected == _PLAYER_FRIENDLY).nonzero()
if(len(player_x) == 0):
return actions.FunctionCall(_NO_OP, [])
player = [int(player_x.mean()), int(player_y.mean())]
points = [player]
closest, min_dist = None, None
other_dest = None
my_dest = None
if ("0" in self.dest_per_marine and "1" in self.dest_per_marine):
if (self.group_id == 0):
my_dest = self.dest_per_marine["0"]
other_dest = self.dest_per_marine["1"]
elif (self.group_id == 1):
other_dest = self.dest_per_marine["0"]
my_dest = self.dest_per_marine["1"]
for p in zip(neutral_x, neutral_y):
if (other_dest):
dist = numpy.linalg.norm(
numpy.array(other_dest) - numpy.array(p))
if (dist < 5):
#print("continue since partner will take care of it ", p)
continue
pp = [p[0], p[1]]
if (pp not in points):
points.append(pp)
dist = numpy.linalg.norm(numpy.array(player) - numpy.array(p))
if not min_dist or dist < min_dist:
closest, min_dist = p, dist
solve_tsp = False
if (my_dest):
dist = numpy.linalg.norm(
numpy.array(player) - numpy.array(my_dest))
if (dist < 2):
solve_tsp = True
if (my_dest is None):
solve_tsp = True
if (len(points) < 2):
solve_tsp = False
if (solve_tsp):
# function for printing best found solution when it is found
from time import clock
init = clock()
def report_sol(obj, s=""):
print("cpu:%g\tobj:%g\ttour:%s" % \
(clock(), obj, s))
#print("points: %s" % points)
n, D = mk_matrix(points, distL2)
# multi-start local search
#print("random start local search:", n)
niter = 50
tour, z = multistart_localsearch(niter, n, D)
#print("best found solution (%d iterations): z = %g" % (niter, z))
#print(tour)
left, right = None, None
for idx in tour:
if (tour[idx] == 0):
if (idx == len(tour) - 1):
#print("optimal next : ", tour[0])
right = points[tour[0]]
left = points[tour[idx - 1]]
elif (idx == 0):
#print("optimal next : ", tour[idx+1])
right = points[tour[idx + 1]]
left = points[tour[len(tour) - 1]]
else:
#print("optimal next : ", tour[idx+1])
right = points[tour[idx + 1]]
left = points[tour[idx - 1]]
left_d = numpy.linalg.norm(
numpy.array(player) - numpy.array(left))
right_d = numpy.linalg.norm(
numpy.array(player) - numpy.array(right))
if (right_d > left_d):
closest = left
else:
closest = right
#print("optimal next :" , closest)
self.dest_per_marine[str(self.group_id)] = closest
#print("dest_per_marine", self.dest_per_marine)
#dest_per_marine {'0': [56, 26], '1': [52, 6]}
if (closest is None):
return actions.FunctionCall(_NO_OP, [])
return actions.FunctionCall(_MOVE_SCREEN, [_NOT_QUEUED, closest])
elif (len(self.group_list) > 0):
player_p = []
for p in zip(player_x, player_y):
if (p not in player_p):
player_p.append(p)
self.group_id = random.randint(0, len(self.group_list) - 1)
return actions.FunctionCall(
_SELECT_CONTROL_GROUP,
[[_CONTROL_GROUP_RECALL], [int(self.group_id)]])
else:
return actions.FunctionCall(_NO_OP, [])