def noop_board_sequence(obs, length):
"""
Simulate the sequence of boards, assuming agents stay unmoved
"""
model = ForwardModel()
# Dummy objects
actions = [constants.Action.Stop.value] * 4 # agents stay unmoved
curr_agents = list() # empty list of Bombers
curr_items = dict() # we never know hidden items
# Prepare initial state
curr_board = obs["board"]
curr_bombs = list()
rows, cols = np.where(obs["bomb_life"] > 0)
for row, col in zip(rows, cols):
bomber = characters.Bomber() # dummy owner of the bomb
position = (row, col)
life = int(obs["bomb_life"][row][col])
blast_strength = int(obs["bomb_blast_strength"][row][col])
moving_direction = None # TODO: this may be known
bomb = characters.Bomb(bomber,
position,
life,
blast_strength,
moving_direction)
curr_bombs.append(bomb)
# overwrite bomb over agent if they overlap
curr_board[position] = constants.Item.Bomb.value
curr_flames = list()
rows, cols = np.where(obs["board"] == constants.Item.Flames.value)
for row, col in zip(rows, cols):
position = (row, col)
life = None # TODO: this may be known
if life is not None:
flame = characters.Flame(position, life)
else:
flame = characters.Flame(position)
curr_flames.append(flame)
# Simulate
list_boards = [curr_board.copy()]
for _ in range(length):
curr_board, _, curr_bombs, _, curr_flames = model.step(actions,
curr_board,
curr_agents,
curr_bombs,
curr_items,
curr_flames)
list_boards.append(curr_board.copy())
return list_boards
def get_initial_game_data(obs, my_id, max_steps=1000):
game_data = GameData()
game_data.board_size = len(obs['board'])
game_data.step_count = obs['step_count'] - 1
game_data.max_steps = max_steps
game_data.game_type = obs['game_type']
game_data.simulation_bomb_life = None
# board
game_data.board = EnvSimulator.get_board(game_data.board_size,
obs['board'])
# items
game_data.items = {}
# agents
game_data.agents = []
for id in [
constants.Item.Agent0.value - 10,
constants.Item.Agent1.value - 10
]:
board_id = id + 10
agent = characters.Bomber(id, game_data.game_type)
agent.set_start_position(
EnvSimulator.get_position(game_data.board, board_id, True))
if (id == my_id):
agent.reset(int(obs['ammo']), board_id in obs['alive'],
int(obs['blast_strength']), bool(obs['can_kick']))
else:
agent.reset(agent.ammo, board_id in obs['alive'],
agent.blast_strength, agent.can_kick)
game_data.agents.append(agent)
# bombs
game_data.bombs = []
bomb_array = EnvSimulator.get_position(game_data.board,
constants.Item.Bomb.value,
False)
if len(bomb_array) > 0:
raise ValueError('Invalid: no bombs allowed in initial state')
# flames
game_data.flames = []
flame_array = EnvSimulator.get_position(game_data.board,
constants.Item.Flames.value,
False)
if len(flame_array) > 0:
raise ValueError('Invalid: no flames allowed in initial state')
# done
game_data.done = forward_model.ForwardModel.get_done(
game_data.agents, game_data.step_count, game_data.max_steps,
game_data.game_type, None)
return game_data
def __init__(self, obs, init=False, bombing_agents={}, board_size=11):
self._game_mode = constants.GameType.FFA
self.move = None
self._board_size = board_size
self._obs = obs
self._my_position = tuple(obs['position'])
self._board = np.array(obs['board'])
self._bomb_life = np.array(self._obs['bomb_life'])
self._teammate = obs['teammate']
self._enemies = [constants.Item(e) for e in obs['enemies']]
self._ammo = int(obs['ammo'])
self.fm = forward_model.ForwardModel()
self.self_agent = self.find_self_agent(self._obs)
agents_id = [
constants.Item.Agent0, constants.Item.Agent1, constants.Item.Agent2,
constants.Item.Agent3
]
self._agents = [
characters.Bomber(aid.value, "FFA") for aid in agents_id
] # remember to modifiy if it is team or radio mode
self.bombing_agents = copy.deepcopy(bombing_agents)
self.score = 0
if init:
self.curr_flames = self.convert_flames(
self._board) # determine by confirming the map
self.curr_bombs = self.convert_bombs(
np.array(obs['bomb_blast_strength']),
np.array(obs['bomb_life']))
self.curr_items = self.convert_items(self._board)
self.curr_agents = self.convert_agents(self._board)
self.last_items = self.curr_items
if (bombing_agents != {}):
self.curr_bombs = self.convert_bombs_two(
np.array(self._obs['bomb_blast_strength']), self._bomb_life,
bombing_agents)
def get_gamedata(gamestate, game_type):
game_data = env_simulator.GameData()
game_data.board_size = gamestate['board_size']
game_data.step_count = gamestate['step_count'] - 1
game_data.max_steps = 800
game_data.game_type = game_type
game_data.simulation_bomb_life = None
# board
game_data.board = gamestate['board']
# items
game_data.items = {}
# agents
game_data.agents = []
for a in gamestate['agents']:
id = a['agent_id']
board_id = id + 10
agent = characters.Bomber(id, game_data.game_type)
agent.set_start_position(get_position(game_data.board, board_id, True))
agent.reset(a['ammo'], a['is_alive'], a['blast_strength'],
a['can_kick'])
game_data.agents.append(agent)
# bombs
game_data.bombs = []
for b in gamestate['bombs']:
bomb = characters.Bomb(**b)
game_data.bombs.append(bomb)
# flames
game_data.flames = []
for f in gamestate['flames']:
flame = characters.Flame(**f)
game_data.flames.append(flame)
# done
game_data.done = forward_model.ForwardModel.get_done(
game_data.agents, game_data.step_count, game_data.max_steps,
game_data.game_type, None)
return game_data
def _get_survivable_actions(self, survivable, obs, curr_bombs, curr_flames):
my_position = obs["position"]
my_blast_strength = obs["blast_strength"]
# is_survivable[action]: whether survivable with action
is_survivable = defaultdict(bool)
x, y = my_position
if (x + 1, y) in survivable[1]:
is_survivable[constants.Action.Down] = True
if (x - 1, y) in survivable[1]:
is_survivable[constants.Action.Up] = True
if (x, y + 1) in survivable[1]:
is_survivable[constants.Action.Right] = True
if (x, y - 1) in survivable[1]:
is_survivable[constants.Action.Left] = True
if (x, y) in survivable[1]:
is_survivable[constants.Action.Stop] = True
# TODO : shoud check the survivability of all agents in one method
# If I have at least one bomb, no bomb in my position,
# and the position is safe
# then consider what happens if I lay a bomb
if all([obs["ammo"] > 0,
obs["bomb_life"][my_position] == 0,
is_survivable[constants.Action.Stop],
sum(is_survivable.values()) > 1]):
board_with_bomb = deepcopy(obs["board"])
curr_bombs_with_bomb = deepcopy(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,
curr_flames,
self._search_range,
my_position)
survivable_with_bomb, prev_bomb \
= self._search_time_expanded_network(list_boards_with_bomb,
my_position)
if my_position in survivable_with_bomb[1]:
is_survivable[constants.Action.Bomb] = True
else:
survivable_with_bomb = None
list_boards_with_bomb = None
return is_survivable, survivable_with_bomb
def _board_sequence(self, board, bombs, flames, length, my_position,
my_action=None, can_kick=False, enemy_mobility=3):
"""
Simulate the sequence of boards, assuming agents stay unmoved
Parameters
----------
board : array
initial board
bombs : list
list of initial bombs
flames : list
list of initial flames
length : int
length of the board sequence to simulate
my_position : tuple
position of my agent
my_action : Action, optional
my action at the first step
can_kick : boolean, optional
whether I can kick
enemy_mobility : int, optional
number of steps where enemies move nondeterministically
Return
------
list_boards : list
list of boards
"""
# Forward model to simulate
model = ForwardModel()
# Prepare initial state
_board = board.copy()
_bombs = deepcopy(bombs)
_flames = deepcopy(flames)
_items = dict() # we never know hidden items
_actions = [constants.Action.Stop.value] * 4
if my_action is not None:
agent = characters.Bomber()
agent.agent_id = board[my_position] - 10
agent.position = my_position
agent.can_kick = can_kick
_agents = [agent]
_actions[agent.agent_id] = my_action
else:
_agents = list()
my_next_position = None
# Get enemy positions to take into account their mobility
rows, cols = np.where(_board > constants.Item.AgentDummy.value)
enemy_positions = [position for position in zip(rows, cols)
if position != my_position]
# List of enemies
enemies = list()
for position in enemy_positions:
agent = characters.Bomber()
agent.agent_id = board[position] - 10
agent.position = position
enemies.append(agent)
_agents = _agents + enemies
# Overwrite bomb over agent if they overlap
for bomb in _bombs:
_board[bomb.position] = constants.Item.Bomb.value
# Simulate
list_boards = [_board.copy()]
for t in range(length):
# Standard simulation step
_board, _agents, _bombs, _, _flames \
= model.step(_actions,
_board,
_agents,
_bombs,
_items,
_flames)
# Overwrite passage over my agent when it has moved to a passage
if t == 0 and len(_agents) > 0:
agent = _agents[0]
my_next_position = agent.position
if all([agent.position != my_position,
_board[agent.position] != constants.Item.Flames.value,
_board[agent.position] != constants.Item.Bomb.value]):
# I did not die and did not stay on a bomb
_board[agent.position] = constants.Item.Passage.value
# Overwrite bomb over agent if they overlap
for bomb in _bombs:
_board[bomb.position] = constants.Item.Bomb.value
# Take into account the nondeterministic mobility of enemies
if t < enemy_mobility:
_enemy_positions = set()
for x, y in enemy_positions:
# for each enemy position in the previous step
for dx, dy in [(0, 0), (1, 0), (-1, 0), (0, 1), (0, -1)]:
# consider the next possible position
next_position = (x + dx, y + dy)
if not self._on_board(next_position):
# ignore if out of board
continue
if any([utility.position_is_passage(_board, next_position),
utility.position_is_powerup(_board, next_position),
(next_position == my_position
and utility.position_is_agent(_board, next_position)
)]):
# possible as a next position
# TODO : what to do with my position
_enemy_positions.add(next_position)
_board[next_position] = constants.Item.AgentDummy.value
enemy_positions = _enemy_positions
_actions = [constants.Action.Stop.value] * 4
_agents = enemies
list_boards.append(_board.copy())
return list_boards, my_next_position
def _get_bombs(self, obs, prev_bomb_life):
"""
Summarize information about bombs
Parameters
----------
obs : dict
pommerman observation
prev_bomb_life : array
remaining life of bombs at the previous step
Return
------
curr_bombs : list
list of bombs
moving_direction : array
array of moving direction of bombs
moving_direction[position] : direction of bomb at position
bomb_life : array
Copy the remaining life of bombs for the next step
"""
# Prepare information about moving bombs
# diff = 0 if no bomb -> no bomb
# diff = 1 if the remaining life of a bomb is decremented
# diff = -9 if no bomb -> bomb
diff = prev_bomb_life - obs["bomb_life"]
moving = (diff != 0) * (diff != 1) * (diff != -9)
# move_from: previous positions of moving bombs
rows, cols = np.where(moving * (diff > 0))
move_from = [position for position in zip(rows, cols)]
# move_to: current positions of moving bombs
rows, cols = np.where(moving * (diff < 0))
move_to = [position for position in zip(rows, cols)]
curr_bombs = list()
rows, cols = np.where(obs["bomb_life"] > 0)
moving_direction = np.full(self.board_shape, None)
for position in zip(rows, cols):
this_bomb_life = obs["bomb_life"][position]
if position in move_to:
# then the bomb is moving, so find the moving direction
for prev_position in move_from:
if prev_bomb_life[prev_position] != this_bomb_life + 1:
# the previous life of the bomb at the previous position
# must be +1 of the life of this bomb
continue
dx = position[0] - prev_position[0]
dy = position[1] - prev_position[1]
if abs(dx) + abs(dy) != 1:
# the previous position must be 1 manhattan distance
# from this position
continue
moving_direction[position] = self._get_direction(prev_position,
position)
# TODO: there might be multiple possibilities of
# where the bomb came from
break
bomb = characters.Bomb(characters.Bomber(), # dummy owner of the bomb
position,
this_bomb_life,
int(obs["bomb_blast_strength"][position]),
moving_direction[position])
curr_bombs.append(bomb)
return curr_bombs, moving_direction, obs["bomb_life"].copy()
def act(self, obs, action_space, info):
#
# Definitions
#
self._search_range = 10
board = obs['board']
my_position = obs["position"] # tuple([x,y]): my position
my_ammo = obs['ammo'] # int: the number of bombs I have
my_blast_strength = obs['blast_strength']
my_enemies = [constants.Item(e) for e in obs['enemies']]
my_teammate = obs["teammate"]
my_kick = obs["can_kick"] # whether I can kick
print("my position", my_position, end="\t")
#
# Understand current situation
#
# List of the set of survivable time-positions at each time
# and preceding positions
survivable_no_move, prev_no_move \
= self._search_time_expanded_network(info["list_boards_no_move"],
my_position)
# Items that can be reached in a survivable manner
reachable_items_no_move, reached_no_move, next_to_items_no_move \
= self._find_reachable_items(info["list_boards_no_move"],
my_position,
survivable_no_move)
# Simulation assuming enemies move
for enemy_mobility in range(3, -1, -1):
# List of boards simulated
list_boards, _ = self._board_sequence(
board,
info["curr_bombs"],
info["curr_flames"],
self._search_range,
my_position,
enemy_mobility=enemy_mobility)
# List of the set of survivable time-positions at each time
# and preceding positions
survivable, prev = self._search_time_expanded_network(
list_boards, my_position)
if len(survivable[1]) > 0:
# Gradually reduce the mobility of enemy, so we have at least one survivable action
break
# Items that can be reached in a survivable manner
reachable_items, reached, next_to_items \
= self._find_reachable_items(list_boards,
my_position,
survivable)
# Survivable actions
is_survivable, survivable_with_bomb \
= self._get_survivable_actions(survivable,
obs,
info["curr_bombs"],
info["curr_flames"])
survivable_actions = [a for a in is_survivable if is_survivable[a]]
# if verbose:
if True:
print("survivable actions are", survivable_actions)
# Positions where we kick a bomb if we move to
if my_kick:
kickable = self._kickable_positions(obs, info["moving_direction"])
else:
kickable = set()
print()
for t in range(0):
print(list_boards[t])
print(survivable[t])
for key in prev[t]:
print(key, prev[t][key])
#
# Choose an action
#
"""
# This is not effective in the current form
if len(survivable_actions) > 1:
# avoid the position if only one position at the following step
# the number of positions that can be reached from the next position
next = defaultdict(set)
next_count = defaultdict(int)
for position in survivable[1]:
next[position] = set([p for p in prev[2] if position in prev[2][p]])
next_count[position] = len(next[position])
print("next count", next_count)
if max(next_count.values()) > 1:
for position in survivable[1]:
if next_count[position] == 1:
risky_action = self._get_direction(my_position, position)
is_survivable[risky_action] = False
survivable_actions = [a for a in is_survivable if is_survivable[a]]
"""
# Do not stay on a bomb if I can
if all([
obs["bomb_life"][my_position] > 0,
len(survivable_actions) > 1,
is_survivable[constants.Action.Stop]
]):
is_survivable[constants.Action.Stop] = False
survivable_actions = [a for a in is_survivable if is_survivable[a]]
if len(survivable_actions) == 0:
print("Must die")
return None
elif len(survivable_actions) == 1:
# move to the position if it is the only survivable position
action = survivable_actions[0]
print("The only survivable action", action)
return action.value
# TODO : shoud check the survivability of all agents in one method
# Place a bomb if
# - it does not significantly reduce my survivability
# - it can reduce the survivability of enemies
consider_bomb = True
if survivable_with_bomb is None:
consider_bomb = False
elif any([len(s) <= 2 for s in survivable_with_bomb[1:]]):
# if not sufficiently survivable all the time after bomb, do not bomb
consider_bomb = False
if consider_bomb:
# place bomb if can reach fog/enemy
if self._can_break(info["list_boards_no_move"][-1], my_position,
my_blast_strength,
[constants.Item.Fog] + my_enemies):
print("Bomb to break fog/enemy", constants.Action.Bomb)
print(info["list_boards_no_move"][-1])
return constants.Action.Bomb.value
for enemy in my_enemies:
# check if the enemy is reachable
if len(reachable_items_no_move[enemy]) == 0:
continue
# can reach the enemy at enemy_position in enemy_time step
enemy_time = reachable_items_no_move[enemy][0][0]
enemy_position = reachable_items_no_move[enemy][0][1:3]
# check if placing a bomb can reduce the survivability
# of the enemy
survivable_before, _ = self._search_time_expanded_network(
info["list_boards_no_move"], enemy_position)
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)
survivable_after, _ \
= self._search_time_expanded_network(list_boards_with_bomb,
enemy_position)
good_before = np.array([len(s) for s in survivable_before])
good_after = np.array([len(s) for s in survivable_after])
# TODO : what are good criteria?
if any(good_after < good_before):
# place a bomb if it makes sense
print("Bomb to kill an enemy", constants.Action.Bomb)
print("before", good_before)
print("after ", good_after)
print([len(s) for s in survivable])
print([len(s) for s in survivable_with_bomb])
return constants.Action.Bomb.value
"""
# find direction towards enemy
positions = set([x[1:3] for x in next_to_items_no_move[enemy]])
for t in range(enemy_time, 1, -1):
_positions = set()
for position in positions:
_positions = _positions.union(prev_no_move[t][position])
positions = _positions.copy()
if enemy_time <= my_blast_strength:
#if True:
positions.add(my_position)
positions_after_bomb = set(survivable[1]).difference(positions)
if positions_after_bomb:
print("Bomb to kill an enemy", enemy, constants.Action.Bomb)
return constants.Action.Bomb.value
"""
# if I can kick, consider placing a bomb to kick
if my_kick and my_position in survivable_with_bomb[3]:
# consdier a sequence of actions: place bomb -> move (action) -> move back (kick)
for action in [
constants.Action.Up, constants.Action.Down,
constants.Action.Left, constants.Action.Right
]:
if not is_survivable[action]:
continue
if action == constants.Action.Up:
# kick direction is down
dx = 1
dy = 0
elif action == constants.Action.Down:
# kick direction is up
dx = -1
dy = 0
elif action == constants.Action.Left:
# kick direction is right
dx = 0
dy = 1
elif action == constants.Action.Right:
# kick direction is left
dx = 0
dy = -1
else:
raise ValueError()
_next_position = (my_position[0] + dx, my_position[1] + dy)
if not self._on_board(_next_position):
continue
else:
next_position = _next_position
# Find where the bomb stops if kicked
for t in range(int(obs["bomb_life"][my_position]) - 2):
if not utility.position_is_passage(
board, next_position):
break
_next_position = (next_position[0] + dx,
next_position[1] + dy)
if not self._on_board(_next_position):
break
else:
next_position = _next_position
if utility.position_is_fog(board, next_position):
print("Bomb to kick into fog", action)
return constants.Action.Bomb.value
elif utility.position_is_enemy(list_boards[t + 2],
next_position,
my_enemies):
print("Bomb to kick towards enemy", action)
return constants.Action.Bomb.value
"""
x0, y0 = my_position
positions_against = [(2*x0-x, 2*y0-y) for (x, y) in positions]
positions_after_bomb = set(survivable[1]).intersection(positions_against)
if positions_after_bomb:
print("Bomb to kick", enemy, constants.Action.Bomb)
return constants.Action.Bomb.value
"""
# kick
if len(kickable) > 0:
while kickable:
# then consider what happens if I kick a bomb
next_position = kickable.pop()
# do not kick a bomb if it will break enemies
if info["moving_direction"][next_position] is None:
# if it is a static bomb
if self._can_break(info["list_boards_no_move"][0],
next_position, my_blast_strength,
my_enemies):
continue
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=3)
survivable_with_kick, prev_kick \
= self._search_time_expanded_network(list_boards_with_kick[1:],
next_position)
if next_position in survivable_with_kick[0]:
print("Kicking", my_action)
return my_action.value
# if on a bomb, consider where to kick in the following step
if obs["bomb_life"][my_position] > 0:
# For each survivable move in the next step,
# check what happens if we kick in the following step.
# If the bomb is kicked into a fog, plan to kick.
# If the bomb is kicked toward an enemy, plan to kick.
# Otherwise, do not plan to kick.
for action in [
constants.Action.Up, constants.Action.Down,
constants.Action.Left, constants.Action.Right
]:
if not is_survivable[action]:
continue
if action == constants.Action.Up:
# kick direction is down
dx = 1
dy = 0
elif action == constants.Action.Down:
# kick direction is up
dx = -1
dy = 0
elif action == constants.Action.Left:
# kick direction is right
dx = 0
dy = 1
elif action == constants.Action.Right:
# kick direction is left
dx = 0
dy = -1
else:
raise ValueError()
_next_position = (my_position[0] + dx, my_position[1] + dy)
if not self._on_board(_next_position):
continue
else:
next_position = _next_position
# Find where the bomb stops if kicked
for t in range(int(obs["bomb_life"][my_position]) - 1):
if not utility.position_is_passage(board, next_position):
break
_next_position = (next_position[0] + dx,
next_position[1] + dy)
if not self._on_board(_next_position):
break
else:
next_position = _next_position
if utility.position_is_fog(board, next_position):
print("Moving to kick into fog", action)
return action.value
elif utility.position_is_enemy(list_boards[t + 2],
next_position, my_enemies):
print("Moving to kick towards enemy", action)
# Move towards an enemy
good_time_positions = set()
for enemy in my_enemies:
good_time_positions = good_time_positions.union(
next_to_items[enemy])
if len(good_time_positions) > 0:
action = self._find_distance_minimizer(my_position,
good_time_positions, prev,
is_survivable)
if action is not None:
print("Moving toward enemy", action)
return action.value
#
# Random action
#
action = random.choice(survivable_actions)
print("Random action", action)
return action.value
def _get_bombs(self, board, bomb_blast_strength, prev_bomb_blast_strength,
bomb_life, prev_bomb_life):
"""
Summarize information about bombs
Parameters
----------
board : array
bomb_blast_strength : array
bomb_life : array
prev_bomb_life : array
remaining life of bombs at the previous step
Return
------
curr_bombs : list
list of bombs
moving_direction : array
array of moving direction of bombs
moving_direction[position] : direction of bomb at position
bomb_life : array
Copy the remaining life of bombs for the next step
"""
# Keep bombs under fog
bomb_positions_under_fog = np.where(
(prev_bomb_life > 1) * (board == constants.Item.Fog.value))
bomb_life[bomb_positions_under_fog] = prev_bomb_life[
bomb_positions_under_fog] - 1
bomb_blast_strength[
bomb_positions_under_fog] = prev_bomb_blast_strength[
bomb_positions_under_fog]
# Prepare information about moving bombs
# diff = 0 if no bomb -> no bomb
# diff = 1 if the remaining life of a bomb is decremented
# diff = -9 if no bomb -> new bomb
diff = prev_bomb_life - bomb_life
moving = (diff != 0) * (diff != 1) * (diff != -9)
# move_from: previous positions of moving bombs
rows, cols = np.where(moving * (diff > 0))
move_from = [position for position in zip(rows, cols)]
# move_to: current positions of moving bombs
rows, cols = np.where(moving * (diff < 0))
move_to = [position for position in zip(rows, cols)]
# TODO : Consider bombs moving into fog
matched_move_from = [False] * len(move_from)
curr_bombs = list()
rows, cols = np.where(bomb_life > 0)
moving_direction = np.full(self.board_shape, None)
for position in zip(rows, cols):
this_bomb_life = bomb_life[position]
if position in move_to:
# then the bomb is moving, so find the moving direction
for i, prev_position in enumerate(move_from):
if prev_bomb_life[prev_position] != this_bomb_life + 1:
# the previous life of the bomb at the previous position
# must be +1 of the life of this bomb
continue
dx = position[0] - prev_position[0]
dy = position[1] - prev_position[1]
if abs(dx) + abs(dy) == 2:
# this can be a moving bomb whose direction is changed by kick
agent_position = (prev_position[0] + dx,
prev_position[1])
if utility.position_is_agent(board, agent_position):
# the agent must have kicked
print("agent must have kicked at", agent_position)
moving_direction[position] = self._get_direction(
agent_position, position)
break
agent_position = (prev_position[0],
prev_position[1] + dy)
if utility.position_is_agent(board, agent_position):
# the agent must have kicked
print("agent must have kicked at", agent_position)
moving_direction[position] = self._get_direction(
agent_position, position)
break
if abs(dx) + abs(dy) != 1:
# the previous position must be 1 manhattan distance
# from this position
continue
moving_direction[position] = self._get_direction(
prev_position, position)
# TODO: there might be multiple possibilities of
# where the bomb came from
matched_move_from[i] = True
break
bomb = characters.Bomb(
characters.Bomber(), # dummy owner of the bomb
position,
this_bomb_life,
int(bomb_blast_strength[position]),
moving_direction[position])
curr_bombs.append(bomb)
return curr_bombs, moving_direction
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
