def get_possible_actions(self, obs, board, pos, ammo, can_kick, bombs):
valid_acts = [0]
x, y = pos
dirX = [-1, 1, 0, 0]
dirY = [ 0, 0, -1, 1]
for k in range(0, len(dirX)):
newX = x + dirX[k]
newY = y + dirY[k]
# print((newX, newY), board.shape)
if newX < board.shape[0] and newY < board.shape[1] and newX >=0 and newY >= 0:
cbom = self.check_bomb((newX, newY), bombs)
if ((board[newX, newY] in [0, 5, 6, 7, 8]) and (not cbom)):
valid_acts.append(k+1)
elif board[newX, newY] in [3] and can_kick:
valid_acts.append(k+1)
#print('contributed to suicide !')
elif board[newX, newY] in [0, 6, 7, 8] and utility.position_is_bomb(bombs, (x,y)):
#print('contributed to death !!!')
valid_acts.append(k+1)
#print('Appending ', k+1, newX, newY, cbom)
if ammo > 0:
valid_acts.append(5)
if len(valid_acts) > 1 and utility.position_is_bomb(bombs, (pos[0], pos[1])) and self.check_bomb((pos[0], pos[1]), bombs):
valid_acts.pop(0)
for i in range(6, len(Actions)):
if self.virtual_actions[i].is_valid(self.cur_state, obs):
#print("appending act ", i)
valid_acts.append(i)
return valid_acts
def get_observation_state(self, board, pos, enemies, bomb_map):
"""
Need just the board layout to decide everything
board -> np.array
pos -> tuple
enemies -> list
"""
def convert_bombs(bomb_map):
'''Flatten outs the bomb array'''
ret = []
locations = np.where(bomb_map > 0)
for r, c in zip(locations[0], locations[1]):
ret.append(crazy_util.dotdict({
'position': (r, c),
'blast_strength': int(bomb_map[(r, c)])
}))
return ret
bombs = convert_bombs(np.array(bomb_map))
has_bomb = False
has_enemy = False
has_wood = False
los_bomb = False
x, y = pos
dirX = [-1,0,1,0]
dirY = [0,1,0,-1]
for k1 in dirX:
for k2 in dirY:
newX = x+k1
newY = y+k2
# print((newX, newY), board.shape)
if newX < board.shape[0] and newY < board.shape[1] and newX >=0 and newY >= 0:
if utility.position_is_bomb(bombs, (newX, newY)):
has_bomb = True
if utility.position_is_wood(board, (newX, newY)):
has_wood = True
if utility.position_is_enemy(board, pos, enemies):
has_enemy = True
for k1 in range(0, board.shape[0]):
if utility.position_is_bomb(bombs, (k1, y)):
los_bomb = True
elif utility.position_is_bomb(bombs, (x, k1)):
los_bomb = True
if utility.position_is_bomb(bombs, (x,y)):
has_bomb = True
if has_bomb:
return 0
elif los_bomb:
return 4
elif has_enemy:
return 1
elif has_wood:
return 2
else:
return 3
def get_observation_state(self, board, pos, enemies, bomb_map, bomb_life,
ammo):
"""
Need just the board layout to decide everything
board -> np.array
pos -> tuple
enemies -> list
"""
bombs = self.convert_bombs(np.array(bomb_map), np.array(bomb_life))
has_bomb = False
has_enemy = False
has_wood = False
los_bomb = False
has_ammo = False
if ammo > 0:
has_ammo = True
x, y = pos
dirX = [-1, 1, 0, 0]
dirY = [0, 0, -1, 1]
for k in range(0, len(dirX)):
newX = x + dirX[k]
newY = y + dirY[k]
# print((newX, newY), board.shape)
if newX < board.shape[0] and newY < board.shape[
1] and newX >= 0 and newY >= 0:
if utility.position_is_bomb(bombs, (newX, newY)):
has_bomb = True
if utility.position_is_wood(board, (newX, newY)):
has_wood = True
if utility.position_is_enemy(board, pos, enemies):
has_enemy = True
los_bomb = self.check_bomb((newX, newY), bombs)
if utility.position_is_bomb(bombs, (x, y)):
has_bomb = True
state = 3
if has_bomb:
state = 0
elif los_bomb:
state = 4
elif has_enemy:
state = 1
elif has_wood:
state = 2
else:
state = 3
if has_ammo:
state = 2 * state
return state
def get_observation_state(self, board, pos, enemies, bomb_map, bomb_life,
ammo, can_kick):
"""
Need just the board layout to decide everything
board -> np.array
pos -> tuple
enemies -> list
"""
bombs = self.convert_bombs(np.array(bomb_map), np.array(bomb_life))
has_bomb = False
has_enemy = False
# is_surrounded = False
is_surrounded = False
los_bomb = False
has_ammo = False
# can kick is also a valid state
if ammo > 0:
has_ammo = True
x, y = pos
dirX = [-1, 1, 0, 0]
dirY = [0, 0, -1, 1]
blocks = 0
for k in range(0, len(dirX)):
newX = x + dirX[k]
newY = y + dirY[k]
# print((newX, newY), board.shape)
if newX < board.shape[0] and newY < board.shape[
1] and newX >= 0 and newY >= 0:
if utility.position_is_bomb(bombs, (newX, newY)):
has_bomb = True
if utility.position_is_rigid(board, (newX, newY)):
# is_surrounded = True
blocks += 1
if utility.position_is_enemy(board, pos, enemies):
has_enemy = True
los_bomb = self.check_bomb((newX, newY), bombs) or los_bomb
if utility.position_is_bomb(bombs, (x, y)) or self.check_bomb(
(x, y), bombs):
has_bomb = True
if blocks > 2:
is_surrounded = True
return State(has_bomb, has_enemy, is_surrounded, los_bomb, has_ammo,
can_kick)
def get_observation_state(self, board, pos, enemies, bomb_map, ammo):
"""
Need just the board layout to decide everything
board -> np.array
pos -> tuple
enemies -> list
"""
bombs = self.convert_bombs(np.array(bomb_map))
has_bomb = False
has_enemy = False
has_wood = False
los_bomb = False
has_ammo = False
if ammo > 0:
has_ammo = True
x, y = pos
dirX = [-1, 1, 0, 0]
dirY = [0, 0, -1, 1]
for k in range(0, len(dirX)):
newX = x + dirX[k]
newY = y + dirY[k]
# print((newX, newY), board.shape)
if newX < board.shape[0] and newY < board.shape[
1] and newX >= 0 and newY >= 0:
if utility.position_is_bomb(bombs, (newX, newY)):
has_bomb = True
if utility.position_is_wood(board, (newX, newY)):
has_wood = True
if utility.position_is_enemy(board, pos, enemies):
has_enemy = True
for bomb in bombs:
if ((abs(newX - bomb['position'][0]) <=
bomb['blast_strength']
and newY == bomb['position'][1])
or (abs(newY - bomb['position'][1]) <=
bomb['blast_strength']
and newX == bomb['position'][0])):
los_bomb = True
if utility.position_is_bomb(bombs, (x, y)):
has_bomb = True
return State(has_bomb, has_enemy, has_wood, los_bomb, has_ammo)
def get_possible_actions(self, board, pos, ammo, can_kick, bombs):
"""
0 : Pass
1 : Up
2 : Down
3 : Left
4 : Right
5 : Bomb
"""
valid_acts = [0]
x, y = pos
dirX = [-1, 1, 0, 0]
dirY = [0, 0, -1, 1]
#print(bombs)
for k in range(0, len(dirX)):
newX = x + dirX[k]
newY = y + dirY[k]
# print((newX, newY), board.shape)
if newX < board.shape[0] and newY < board.shape[
1] and newX >= 0 and newY >= 0:
cbom = self.check_bomb((newX, newY), bombs)
if ((board[newX, newY] in [0, 5, 6, 7, 8]) and (not cbom)):
valid_acts.append(k + 1)
elif board[newX, newY] in [3] and can_kick:
valid_acts.append(k + 1)
#print('contributed to suicide !')
elif board[newX, newY] in [0, 6, 7, 8
] and utility.position_is_bomb(
bombs, (x, y)):
#print('contributed to death !!!')
valid_acts.append(k + 1)
#print('Appending ', k+1, newX, newY, cbom)
if ammo > 0:
valid_acts.append(5)
if len(valid_acts) > 1 and utility.position_is_bomb(
bombs, (pos[0], pos[1])) and self.check_bomb(
(pos[0], pos[1]), bombs):
valid_acts.pop(0)
return valid_acts
def get_observation_state(self, obs):
"""
Need just the board layout to decide everything
board -> np.array
pos -> tuple
enemies -> list
Interesting keys:
obs['board'],
obs['position'],
obs['teammate'],
obs['enemies'],
obs['bomb_blast_strength'],
obs['bomb_life'],
obs['ammo'],
obs['can_kick']
"""
board = obs["board"]
bombs = self.convert_bombs(np.array(obs["bomb_blast_strength"]), np.array(obs["bomb_life"]))
d = collections.OrderedDict({
"bomb_nearby": Proximity.NONE,
"enemy_nearby": Proximity.NONE,
"is_surrounded": False,
"los_bomb": False,
"ammo": 3 if obs['ammo'] > 3 else obs['ammo'],
"can_kick": obs['can_kick'],
"blast_strength": BlastStrength.LOW if obs['blast_strength'] <= 2 else BlastStrength.HIGH ,
"enemies_alive": len(list(filter(lambda enemy: enemy.value in obs['alive'], obs['enemies']))),
"nearby_enemy_has_bomb": False,
"nearby_enemy_can_kick": False,
"next_to_wood": False
})
x, y = obs['position']
nearby_enemy_id = None
for del_x in range(-2, 3):
for del_y in range(-2, 3):
newX = x + del_x
newY = y + del_y
immediate_zone = abs(del_x) <= 1 and abs(del_y) <= 1
if newX < board.shape[0] and newY < board.shape[1] and newX >= 0 and newY >= 0:
if utility.position_is_bomb(bombs, (newX, newY)):
d['bomb_nearby'] = Proximity.IMMEDIATE if immediate_zone else Proximity.CLOSE
if immediate_zone and obs["board"][newX, newY] == 2:
d["next_to_wood"] = True
if utility.position_is_enemy(obs['board'], (newX, newY), obs['enemies']):
nearby_enemy_id = obs['board'][newX, newY]
d['enemy_nearby'] = Proximity.IMMEDIATE if immediate_zone else Proximity.CLOSE
d['los_bomb'] = self.check_bomb((newX, newY), bombs) or d['los_bomb']
if utility.position_is_bomb(bombs, (x,y)) or self.check_bomb((x,y), bombs): # TODO why two conditions?
d["bomb_nearby"] = Proximity.IMMEDIATE
d["is_surrounded"] = ep.is_pos_surrounded(obs["board"], obs["position"], self.agent_value)
#print(d["is_surrounded"])
if nearby_enemy_id and self.enemy_info:
#print(self.enemy_info)
enemy_object = self.enemy_info[nearby_enemy_id - 10] # 10, 11, 12, 13 index, one assumes
d["nearby_enemy_has_bomb"] = enemy_object['ammo'] > 0
d["nearby_enemy_can_kick"] = enemy_object['can_kick'] > 0
return AliveState(**d)
def act(self, obs, action_space, info):
#
# Definitions
#
board = obs['board']
my_position = obs["position"] # tuple([x,y]): my position
my_kick = obs["can_kick"] # whether I can kick
my_enemies = [constants.Item(e) for e in obs['enemies']]
my_teammate = obs["teammate"]
# List of the set of survivable time-positions at each time
# and preceding positions
survivable, prev, succ, _ \
= self._search_time_expanded_network(info["list_boards_no_move"],
my_position)
# Survivable actions
is_survivable, survivable_with_bomb \
= self._get_survivable_actions(survivable,
obs,
info["curr_bombs"],
info["curr_flames"],
enemy_mobility=0)
survivable_actions = [a for a in is_survivable if is_survivable[a]]
n_survivable = dict()
kick_actions = list()
if my_kick:
# Positions where we kick a bomb if we move to
kickable = self._kickable_positions(obs, info["moving_direction"])
for next_position in kickable:
# consider what happens if I kick a bomb
my_action = self._get_direction(my_position, next_position)
# do not kick into fog
dx = next_position[0] - my_position[0]
dy = next_position[1] - my_position[1]
position = next_position
is_fog = False
while self._on_board(position):
if utility.position_is_fog(board, position):
is_fog = True
break
position = (position[0] + dx, position[1] + dy)
if is_fog:
continue
list_boards_with_kick, next_position \
= self._board_sequence(obs["board"],
info["curr_bombs"],
info["curr_flames"],
self._search_range,
my_position,
my_action=my_action,
can_kick=True,
enemy_mobility=0)
#print(list_boards_with_kick)
survivable_with_kick, prev_kick, succ_kick, _ \
= self._search_time_expanded_network(list_boards_with_kick[1:],
next_position)
if next_position in survivable_with_kick[0]:
survivable_actions.append(my_action)
is_survivable[my_action] = True
n_survivable[my_action] = [1] + [
len(s) for s in survivable_with_kick[1:]
]
kick_actions.append(my_action)
else:
kickable = set()
x, y = my_position
for action in survivable_actions:
# for each survivable action, check the survivability
if action == constants.Action.Bomb:
n_survivable[action] = [
len(s) for s in survivable_with_bomb[1:]
]
continue
if action == constants.Action.Up:
dx = -1
dy = 0
elif action == constants.Action.Down:
dx = 1
dy = 0
elif action == constants.Action.Left:
dx = 0
dy = -1
elif action == constants.Action.Right:
dx = 0
dy = 1
elif action == constants.Action.Stop:
dx = 0
dy = 0
else:
raise ValueError()
next_position = (x + dx, y + dy)
n_survivable[action], _info = self._count_survivable(
succ, 1, next_position)
most_survivable_action = None
if survivable_actions:
survivable_score = dict()
for action in n_survivable:
#survivable_score[action] = sum([-n**(-5) for n in n_survivable[action]])
survivable_score[action] = sum(
[n for n in n_survivable[action]])
if verbose:
print(action, survivable_score[action],
n_survivable[action])
best_survivable_score = max(survivable_score.values())
random.shuffle(survivable_actions)
for action in survivable_actions:
if survivable_score[action] == best_survivable_score:
most_survivable_action = action
break
if most_survivable_action is not None:
print("Most survivable action", most_survivable_action)
return most_survivable_action.value
# kick if possible
if my_kick:
kickable = self._kickable_positions(obs, info["moving_direction"])
else:
kickable = set()
print("Kickable", my_kick, kickable)
while kickable:
next_position = kickable.pop()
action = self._get_direction(my_position, next_position)
print("Must kick to survive", action)
return action.value
# move towards a teammate if she is blocking
for action in [
constants.Action.Right, constants.Action.Left,
constants.Action.Down, constants.Action.Up
]:
next_position = utility.get_next_position(my_position, action)
if not self._on_board(next_position):
continue
if utility._position_is_item(board, next_position, my_teammate):
print("Must move to teammate to survive", action)
return action.value
# move towards an enemy
for action in [
constants.Action.Right, constants.Action.Left,
constants.Action.Down, constants.Action.Up
]:
next_position = utility.get_next_position(my_position, action)
if not self._on_board(next_position):
continue
if utility.position_is_enemy(board, next_position, my_enemies):
print("Must move to enemy to survive", action)
return action.value
# move towards anywhere besides ridid
for action in [
constants.Action.Right, constants.Action.Left,
constants.Action.Down, constants.Action.Up
]:
next_position = utility.get_next_position(my_position, action)
if not self._on_board(next_position):
continue
if utility.position_is_rigid(board, next_position):
continue
if utility.position_is_wood(board, next_position):
continue
if utility.position_is_bomb(info["curr_bombs"], next_position):
continue
print("Try moving to survive", action)
return action.value
action = constants.Action.Stop
print("Must die", action)
return action
def act(self, obs, action_space, info):
#
# Definitions
#
board = obs['board']
my_position = obs["position"] # tuple([x,y]): my position
my_kick = obs["can_kick"] # whether I can kick
my_enemies = [constants.Item(e) for e in obs['enemies']]
my_teammate = obs["teammate"]
my_ammo = obs['ammo'] # int: the number of bombs I have
my_blast_strength = obs['blast_strength']
enemy_position = dict()
for enemy in my_enemies:
positions = np.argwhere(board == enemy.value)
if len(positions) == 0:
continue
enemy_position[enemy] = tuple(positions[0])
survivable_steps = defaultdict(int)
#
# survivable tree in standard case
#
list_boards_no_kick = deepcopy(info["list_boards_no_move"])
# remove myself
if obs["bomb_blast_strength"][my_position]:
for b in list_boards_no_kick:
if utility.position_is_agent(b, my_position):
b[my_position] = constants.Item.Bomb.value
else:
for b in list_boards_no_kick:
if utility.position_is_agent(b, my_position):
b[my_position] = constants.Item.Passage.value
my_survivable, my_prev, my_succ, my_survivable_with_enemy \
= self._get_survivable_with_enemy(list_boards_no_kick, my_position, enemy_position)
life = defaultdict(int)
for t in range(self._search_range, 0, -1):
for position in my_survivable_with_enemy[t]:
if not life[(t, ) + position]:
life[(t, ) + position] = t
for prev_position in my_prev[t][position]:
life[(t - 1, ) + prev_position] = max([
life[(t - 1, ) + prev_position], life[(t, ) + position]
])
for next_position in my_survivable[1]:
my_action = self._get_direction(my_position, next_position)
survivable_steps[my_action] = life[(1, ) + next_position]
#
# survivable tree if I lay bomb
#
if all([obs["ammo"] > 0, obs["bomb_life"][my_position] == 0]):
# if I can lay a bomb
board_with_bomb = deepcopy(obs["board"])
curr_bombs_with_bomb = deepcopy(info["curr_bombs"])
# lay a bomb
board_with_bomb[my_position] = constants.Item.Bomb.value
bomb = characters.Bomb(
characters.Bomber(), # dummy owner of the bomb
my_position,
constants.DEFAULT_BOMB_LIFE,
my_blast_strength,
None)
curr_bombs_with_bomb.append(bomb)
list_boards_with_bomb, _ \
= self._board_sequence(board_with_bomb,
curr_bombs_with_bomb,
info["curr_flames"],
self._search_range,
my_position,
enemy_mobility=0)
my_survivable_with_bomb, my_prev_with_bomb, my_succ_with_bomb, my_survivable_with_bomb_enemy \
= self._get_survivable_with_enemy(list_boards_with_bomb, my_position, enemy_position)
life = defaultdict(int)
for t in range(self._search_range, 0, -1):
for position in my_survivable_with_bomb_enemy[t]:
if not life[(t, ) + position]:
life[(t, ) + position] = t
for prev_position in my_prev_with_bomb[t][position]:
life[(t - 1, ) + prev_position] = max([
life[(t - 1, ) + prev_position],
life[(t, ) + position]
])
survivable_steps[constants.Action.Bomb] = life[(1, ) + my_position]
print("survivable steps")
print(survivable_steps)
if survivable_steps:
values = np.array(list(survivable_steps.values()))
print(values)
best_index = np.where(values == np.max(values))
best_actions = np.array(list(survivable_steps.keys()))[best_index]
best_action = random.choice(best_actions)
print("Most survivable action", best_action)
return best_action.value
else:
print("No actions: stop")
return constants.Action.Stop.value
#
# survivable tree if I kick
#
if my_kick:
# Positions where I kick a bomb if I move to
kickable, more_kickable = self._kickable_positions(
obs, info["moving_direction"])
for next_position in set.union(*[kickable, more_kickable]):
# consider what happens if I kick a bomb
my_action = self._get_direction(my_position, next_position)
list_boards_with_kick, next_position \
= self._board_sequence(obs["board"],
info["curr_bombs"],
info["curr_flames"],
self._search_range,
my_position,
my_action=my_action,
can_kick=True,
enemy_mobility=0)
my_survivable_with_kick[next_position], my_prev_with_kick[next_position], my_succ_with_bomb[next_position], my_survivable_with_kick_enemy[next_position] \
= self._get_survivable_with_enemy(list_boards_with_kick[1:], next_position, enemy_position)
survivable_with_kick, prev_kick, succ_kick, _ \
= self._search_time_expanded_network(list_boards_with_kick[1:],
next_position)
# Survivable actions
is_survivable, survivable_with_bomb \
= self._get_survivable_actions(my_survivable,
obs,
info["curr_bombs"],
info["curr_flames"],
enemy_mobility=0)
survivable_actions = [a for a in is_survivable if is_survivable[a]]
n_survivable = dict()
kick_actions = list()
if my_kick:
# Positions where we kick a bomb if we move to
kickable = self._kickable_positions(obs, info["moving_direction"])
for next_position in kickable:
# consider what happens if I kick a bomb
my_action = self._get_direction(my_position, next_position)
list_boards_with_kick, next_position \
= self._board_sequence(obs["board"],
info["curr_bombs"],
info["curr_flames"],
self._search_range,
my_position,
my_action=my_action,
can_kick=True,
enemy_mobility=0)
#print(list_boards_with_kick)
survivable_with_kick, prev_kick, succ_kick, _ \
= self._search_time_expanded_network(list_boards_with_kick[1:],
next_position)
if next_position in survivable_with_kick[0]:
survivable_actions.append(my_action)
is_survivable[my_action] = True
n_survivable[my_action] = [1] + [
len(s) for s in survivable_with_kick[1:]
]
kick_actions.append(my_action)
else:
kickable = set()
x, y = my_position
for action in survivable_actions:
# for each survivable action, check the survivability
if action == constants.Action.Bomb:
n_survivable[action] = [
len(s) for s in survivable_with_bomb[1:]
]
continue
if action == constants.Action.Up:
dx = -1
dy = 0
elif action == constants.Action.Down:
dx = 1
dy = 0
elif action == constants.Action.Left:
dx = 0
dy = -1
elif action == constants.Action.Right:
dx = 0
dy = 1
elif action == constants.Action.Stop:
dx = 0
dy = 0
else:
raise ValueError()
next_position = (x + dx, y + dy)
n_survivable[action], _info = self._count_survivable(
my_succ, 1, next_position)
most_survivable_action = None
if survivable_actions:
survivable_score = dict()
for action in n_survivable:
#survivable_score[action] = sum([-n**(-5) for n in n_survivable[action]])
survivable_score[action] = sum(
[n for n in n_survivable[action]])
if verbose:
print(action, survivable_score[action],
n_survivable[action])
best_survivable_score = max(survivable_score.values())
random.shuffle(survivable_actions)
for action in survivable_actions:
if survivable_score[action] == best_survivable_score:
most_survivable_action = action
break
if most_survivable_action is not None:
print("Most survivable action", most_survivable_action)
return most_survivable_action.value
# kick if possible
if my_kick:
kickable = self._kickable_positions(obs, info["moving_direction"])
else:
kickable = set()
print("Kickable", my_kick, kickable)
while kickable:
next_position = kickable.pop()
action = self._get_direction(my_position, next_position)
print("Must kick to survive", action)
return action.value
# move towards a teammate if she is blocking
for action in [
constants.Action.Right, constants.Action.Left,
constants.Action.Down, constants.Action.Up
]:
next_position = utility.get_next_position(my_position, action)
if not self._on_board(next_position):
continue
if utility._position_is_item(board, next_position, my_teammate):
print("Must move to teammate to survive", action)
return action.value
# move towards an enemy
for action in [
constants.Action.Right, constants.Action.Left,
constants.Action.Down, constants.Action.Up
]:
next_position = utility.get_next_position(my_position, action)
if not self._on_board(next_position):
continue
if utility.position_is_enemy(board, next_position, my_enemies):
print("Must move to enemy to survive", action)
return action.value
# move towards anywhere besides ridid
for action in [
constants.Action.Right, constants.Action.Left,
constants.Action.Down, constants.Action.Up
]:
next_position = utility.get_next_position(my_position, action)
if not self._on_board(next_position):
continue
if utility.position_is_rigid(board, next_position):
continue
if utility.position_is_wood(board, next_position):
continue
if utility.position_is_bomb(info["curr_bombs"], next_position):
continue
print("Try moving to survive", action)
return action.value
action = constants.Action.Stop
print("Must die", action)
return action
