def __call__(self, game: Game):
if len(game.state.reports) < self.last_report_idx:
self.last_report_idx = 0
r = 0.0
own_team = game.active_team
opp_team = game.get_opp_team(own_team)
for outcome in game.state.reports[self.last_report_idx:]:
team = None
if outcome.player is not None:
team = outcome.player.team
elif outcome.team is not None:
team = outcome.team
if team == own_team and outcome.outcome_type in A2C_Reward.rewards_own:
r += A2C_Reward.rewards_own[outcome.outcome_type]
if team == opp_team and outcome.outcome_type in A2C_Reward.rewards_opp:
r += A2C_Reward.rewards_opp[outcome.outcome_type]
self.last_report_idx = len(game.state.reports)
ball_carrier = game.get_ball_carrier()
if ball_carrier is not None:
if self.last_ball_team is own_team and ball_carrier.team is own_team:
ball_progress = self.last_ball_x - ball_carrier.position.x
if own_team is game.state.away_team:
ball_progress *= -1 # End zone at max x coordinate
r += A2C_Reward.ball_progression_reward * ball_progress
self.last_ball_team = ball_carrier.team
self.last_ball_x = ball_carrier.position.x
else:
self.last_ball_team = None
self.last_ball_x = None
return r
def __init__(self, game: botbowl.Game):
self.actions = []
self.priority_actions = []
for action_choice in game.get_available_actions():
if action_choice.action_type not in {
ActionType.START_MOVE, ActionType.MOVE,
ActionType.END_PLAYER_TURN
}:
continue
if action_choice.action_type == ActionType.MOVE:
prio_move_square = get_priority_move_square(
action_choice, game)
if prio_move_square is not None:
self.priority_actions.append(
Action(action_choice.action_type,
position=prio_move_square))
self.actions.extend(convert_to_actions(action_choice))
if len(self.actions) == 0:
self.actions.extend(
collapse(
convert_to_actions(action_choice)
for action_choice in game.get_available_actions()))
assert len(self.actions) > 0
shuffle(self.actions)
def scripted_action(game: botbowl.Game) -> Optional[Action]:
available_action_types = {
action_choice.action_type
for action_choice in game.get_available_actions()
}
for at in use_directly_action_types:
if at in available_action_types:
return Action(at)
if ActionType.PLACE_BALL in available_action_types:
x = game.arena.width // 4
if game.active_team is game.state.away_team:
x *= 3
y = game.arena.height // 2
return Action(ActionType.PLACE_BALL, position=botbowl.Square(x, y))
proc = game.get_procedure()
if type(proc) is Setup:
if game.is_setup_legal(game.active_team):
return Action(ActionType.END_SETUP)
for at in available_action_types:
if at not in {ActionType.END_SETUP, ActionType.PLACE_PLAYER}:
return Action(at)
return None
def path_to_move_actions(game: botbowl.Game,
player: botbowl.Player,
path: Path,
do_assertions=True) -> List[Action]:
"""
This function converts a path into a list of actions corresponding to that path.
If you provide a handoff, foul or blitz path, then you have to manally set action type.
:param game:
:param player: player to move
:param path: a path as returned by the pathfinding algorithms
:param do_assertions: if False, it turns off the validation, can be helpful when the GameState will change before
this path is executed.
:returns: List of actions corresponding to 'path'.
"""
if path.block_dice is not None:
action_type = ActionType.BLOCK
elif path.handoff_roll is not None:
action_type = ActionType.HANDOFF
elif path.foul_roll is not None:
action_type = ActionType.FOUL
else:
action_type = ActionType.MOVE
active_team = game.state.available_actions[0].team
player_at_target = game.get_player_at(path.get_last_step())
if do_assertions:
if action_type is ActionType.MOVE:
assert player_at_target is None or player_at_target is game.get_active_player(
)
elif action_type is ActionType.BLOCK:
assert game.get_opp_team(active_team) is player_at_target.team
assert player_at_target.state.up
elif action_type is ActionType.FOUL:
assert game.get_opp_team(active_team) is player_at_target.team
assert not player_at_target.state.up
elif action_type is ActionType.HANDOFF:
assert active_team is player_at_target.team
assert player_at_target.state.up
else:
raise Exception(f"Unregonized action type {action_type}")
final_action = Action(action_type, position=path.get_last_step())
if game._is_action_allowed(final_action):
return [final_action]
else:
actions = []
if not player.state.up and path.steps[0] == player.position:
actions.append(Action(ActionType.STAND_UP, player=player))
actions.extend(
Action(ActionType.MOVE, position=sq)
for sq in path.steps[1:-1])
else:
actions.extend(
Action(ActionType.MOVE, position=sq) for sq in path.steps[:-1])
actions.append(final_action)
return actions
def __init__(self, game: botbowl.Game, parent: Optional['Node']):
self.step_nbr = game.get_step()
self.parent = parent
self.children = []
self.change_log = game.trajectory.action_log[
parent.step_nbr:] if parent is not None else []
self.top_proc = str(game.get_procedure())
assert parent is None or len(self.change_log) > 0
def hash_game_state(game: botbowl.Game) -> str:
s = ""
s += "h" if game.active_team is game.state.home_team else "a"
s += str(game.state.round)
s += str(game.active_team.state.turn)
s += type(game.get_procedure()).__name__
s += f"{hash(game.get_ball_position())}-"
s += " ".join(
str(hash(p.position)) for p in game.get_players_on_pitch()) + "-"
return s
def expand_throw_in(game: botbowl.Game, parent: Node) -> Node:
# noinspection PyTypeChecker
active_proc: procedures.ThrowIn = game.get_procedure()
assert type(active_proc) is procedures.ThrowIn
d6_fixes = [3, 4] if game.arena.height > 7 else [1, 2]
with only_fixed_rolls(game, d3=[2], d6=d6_fixes):
game.step()
assert active_proc is not game.get_procedure()
return expand_none_action(game, parent)
def expand_pickup(game: botbowl.Game, parent: Node) -> Node:
# noinspection PyTypeChecker
active_proc: procedures.Pickup = game.get_procedure()
assert type(active_proc) is procedures.Pickup
assert active_proc.roll is None
probability_success = game.get_pickup_prob(active_proc.player,
active_proc.ball.position)
new_parent = ChanceNode(game, parent)
# SUCCESS SCENARIO
with only_fixed_rolls(game, d6=[6]):
game.step()
success_node = expand_none_action(game, new_parent, pickup_handled=True)
new_parent.connect_child(success_node, probability_success)
assert game.get_step() == new_parent.step_nbr
# FAILURE SCENARIO
assert not active_proc.player.has_skill(
Skill.SURE_HANDS), 'cant handle sure hands yet'
with only_fixed_rolls(game, d6=[1]):
game.step()
fail_node = expand_none_action(game, new_parent, pickup_handled=True)
new_parent.connect_child(fail_node, 1 - probability_success)
assert game.get_step() == new_parent.step_nbr
return new_parent
def expand_none_action(game: botbowl.Game,
parent: Node,
moving_handled=False,
pickup_handled=False) -> Node:
"""
:param game: the game state is changed during expansion but restored to state of argument 'parent'
:param parent: shall represent the current state of argument 'game'. game state is restored to parent.step_nbr
:param moving_handled:
:param pickup_handled:
:returns: A subclass of Node:
- ChanceNode in a nestled structure with multiple ActionNode as leaf nodes.
- ActionNode if only one possible outcome.
param game is changed but restored to initial state af
Called recursively.
"""
while len(game.state.available_actions) == 0:
proc = game.get_procedure()
expand_func = get_expanding_function(proc, moving_handled,
pickup_handled)
if expand_func is not None:
assert len(botbowl.D6.FixedRolls) == 0
return_node = expand_func(game, parent)
assert len(botbowl.D6.FixedRolls) == 0
game.revert(parent.step_nbr)
return return_node
with only_fixed_rolls(game):
game.step()
action_node = ActionNode(game, parent)
game.revert(parent.step_nbr)
assert parent.step_nbr == game.get_step()
return action_node
def __call__(
self, game: botbowl.Game
) -> Tuple[float, np.ndarray, List[botbowl.Action]]:
action = scripted_action(game)
if action is not None:
return 0.0, np.array([1.0]), [action]
actions: MockPolicy.ActionProbList = []
if self.end_setup:
self.end_setup = False
return 0.0, np.array([1.0]), [Action(ActionType.END_SETUP)]
for action_choice in game.get_available_actions():
action_type = action_choice.action_type
if action_type in self.convert_function:
actions.extend(self.convert_function[action_type](
game, action_choice))
elif action_type in self.positional_types:
if len(action_choice.positions) > 0:
positions = action_choice.positions
else:
positions = [p.position for p in action_choice.players]
for pos in positions:
actions.append((Action(action_type, position=pos), 1))
elif action_choice.action_type in self.simple_types:
actions.append((Action(action_type), 1))
elif type(game.get_procedure()) is Setup and action_type not in {
ActionType.END_SETUP, ActionType.PLACE_PLAYER
}:
actions.append((Action(action_type), 1))
self.end_setup = True
if len(actions) == 0 and ActionType.END_PLAYER_TURN in [
ac.action_type for ac in game.state.available_actions
]:
actions.append((Action(ActionType.END_PLAYER_TURN), 1))
action_objects, probabilities = zip(*actions)
probabilities = np.array(probabilities, dtype=np.float)
probabilities += 0.0001 * np.random.random(len(probabilities))
probabilities /= sum(probabilities)
return 0.0, probabilities, action_objects
def expand_injury(game: botbowl.Game, parent: Node) -> Node:
# noinspection PyTypeChecker
proc: procedures.Injury = game.get_procedure()
assert not proc.foul
if proc.in_crowd:
with only_fixed_rolls(game, d6=[5, 4]): # straight to KO
game.step()
return expand_none_action(game, parent)
p_removal = accumulated_prob_2d_roll[8]
new_parent = ChanceNode(game, parent)
expand_with_fixes(game, new_parent, p_removal, d6=[5, 4]) # KO
expand_with_fixes(game, new_parent, 1 - p_removal, d6=[1, 1]) # Stun
return new_parent
def a2c_scripted_actions(game: Game):
proc_type = type(game.get_procedure())
if proc_type is procedure.Block:
# noinspection PyTypeChecker
return MyScriptedBot.block(self=None, game=game)
return None
def start_handoff_actions(self, game: botbowl.Game,
action_choice) -> ActionProbList:
ball = game.get_ball()
if ball is not None and ball.is_carried:
if ball.position in action_choice.positions:
return [(Action(ActionType.START_HANDOFF,
position=[ball.position]), 1)]
return []
def expand_action(game: botbowl.Game, action: botbowl.Action,
parent: ActionNode) -> Node:
"""
:param game: game object used for calculations. Will be reverted to original state.
:param action: action to be evaluated.
:param parent: parent node
:returns - list of tuples containing (Steps, probability) for each possible outcome.
- probabilities sums to 1.0
Not called recursively
"""
# noinspection PyProtectedMember
assert game._is_action_allowed(action)
assert game.trajectory.enabled
game.config.fast_mode = False
with only_fixed_rolls(game):
game.step(action)
return expand_none_action(game, parent)
def get_filtered_available_actions(game: botbowl.Game):
allowed_avail_actions = []
disallowed = {
ActionType.START_FOUL, ActionType.START_PASS, ActionType.START_HANDOFF
}
for action_choice in game.get_available_actions():
if action_choice.action_type in disallowed:
continue
allowed_avail_actions.append(action_choice)
return allowed_avail_actions
def expand_catch(game: botbowl.Game, parent: Node) -> Node:
# noinspection PyTypeChecker
proc: procedures.Catch = game.get_procedure()
assert type(proc) is procedures.Catch
p_catch = game.get_catch_prob(proc.player,
accurate=proc.accurate,
handoff=proc.handoff)
new_parent = ChanceNode(game, parent)
# Success scenario
expand_with_fixes(game, new_parent, p_catch, d6=[6])
# Failure scenario
if proc.player.has_skill(Skill.CATCH):
expand_with_fixes(game, new_parent, 1 - p_catch, d6=[1, 1])
else:
expand_with_fixes(game, new_parent, 1 - p_catch, d6=[1])
return new_parent
def expand_armor(game: botbowl.Game, parent: Node) -> Node:
# noinspection PyTypeChecker
proc: procedures.Armor = game.get_procedure()
assert not proc.foul
p_armorbreak = accumulated_prob_2d_roll[proc.player.get_av() + 1]
new_parent = ChanceNode(game, parent)
expand_with_fixes(game, new_parent, p_armorbreak, d6=[6,
6]) # Armor broken
expand_with_fixes(game, new_parent, 1 - p_armorbreak,
d6=[1, 1]) # Armor not broken
return new_parent
def act(self, game: Game):
if self.tree is None:
self.tree = SearchTree(deepcopy(game))
else:
self.tree.set_new_root(game)
time_left = game.get_seconds_left()
if time_left < 10:
assert False
# expand the tree
start_time = perf_counter()
while perf_counter() - start_time < 10:
pass
def expand_bounce(game: botbowl.Game, parent: Node) -> Node:
# noinspection PyTypeChecker
active_proc: procedures.Bounce = game.get_procedure()
assert type(active_proc) is procedures.Bounce
new_parent = ChanceNode(game, parent)
ball_pos = active_proc.piece.position
active_player = game.get_active_player()
adj_squares = game.get_adjacent_squares(ball_pos, occupied=False)
sq_to_num_tz = {
sq: game.num_tackle_zones_at(active_player, sq)
for sq in adj_squares
}
num_tz_to_sq = {}
for sq, num_tz in sq_to_num_tz.items():
num_tz_to_sq.setdefault(num_tz, []).append(sq)
for num_tz, count in Counter(sq_to_num_tz.values()).items():
possible_squares = num_tz_to_sq[num_tz]
square = np.random.choice(possible_squares, 1)[0]
assert type(square) == botbowl.Square
delta_x = square.x - ball_pos.x
delta_y = square.y - ball_pos.y
roll = botbowl.D8.d8_from_xy[(delta_x, delta_y)]
with only_fixed_rolls(game, d8=[roll]):
game.step()
new_node = expand_none_action(game, new_parent)
new_parent.connect_child(new_node, prob=count / 8)
assert game.get_step() == new_parent.step_nbr
sum_prob = sum(new_parent.child_probability)
new_parent.child_probability = [
prob / sum_prob for prob in new_parent.child_probability
]
assert sum(new_parent.child_probability) == approx(1.0, abs=1e-9)
assert game.get_step() == new_parent.step_nbr
return new_parent
def handle_ko_wakeup(game: botbowl.Game, parent: Node) -> Node:
# noinspection PyTypeChecker
active_proc: procedures.PreKickoff = game.get_procedure()
assert type(active_proc) is procedures.PreKickoff
d6_fixes = [1] * len(game.get_knocked_out(active_proc.team))
with only_fixed_rolls(game, d6=d6_fixes):
while active_proc is game.get_procedure():
game.step()
assert active_proc is not game.get_procedure()
return expand_none_action(game, parent)
def get_heuristic(game: botbowl.Game) -> HeuristicVector:
"""
Heuristic based on game state, calculated from home teams perspective
zero sum, meaning away team's heuristic is negative of home team's heuristic
:returns: array with different heuristics, multiply it with
"""
score, tv_on_pitch, ball_position, ball_carried, ball_marked = 0.0, 0.0, 0.0, 0.0, 0.0
home = game.state.home_team
away = game.state.away_team
score += home.state.score - away.state.score
tv_on_pitch += sum(p.role.cost
for p in game.get_players_on_pitch(team=home))
tv_on_pitch -= sum(p.role.cost
for p in game.get_players_on_pitch(team=away))
tv_on_pitch /= 50000.0 # normalized to cost of lineman
ball = game.get_ball()
if ball is not None and ball.position is not None:
ball_position -= ball.position.x # negative because home team scores at x = 0
home_marking_ball = len(
game.get_adjacent_players(ball.position, team=home,
standing=True)) > 0
away_marking_ball = len(
game.get_adjacent_players(ball.position, team=away,
standing=True)) > 0
if ball.is_carried:
ball_carrier = game.get_player_at(ball.position)
if ball_carrier.team == home:
ball_carried += 1
ball_marked -= away_marking_ball
elif ball_carrier.team == away:
ball_carried -= 1
ball_marked += home_marking_ball
else:
ball_marked += (home_marking_ball - away_marking_ball)
return HeuristicVector(score=score,
tv_on_pitch=tv_on_pitch,
ball_position=ball_position,
ball_carried=ball_carried,
ball_marked=ball_marked)
def move_actions(self, game: botbowl.Game,
action_choice) -> ActionProbList:
if game.get_player_action_type(
) in self.player_actiontypes_without_move_actions:
return []
action_probs = []
player = game.get_active_player()
ball_carrier = game.get_ball_carrier()
is_ball_carrier = player is game.get_ball_carrier()
if player.state.moves > 0 and not is_ball_carrier:
return []
elif is_ball_carrier and player.state.moves > 0:
action_probs.append((Action(ActionType.END_PLAYER_TURN), 1))
ball = game.get_ball()
ball_on_floor_pos = game.get_ball(
).position if not ball.is_carried else None
opp_ball_carrier = ball_carrier if ball_carrier is not None and ball_carrier.team is not game.active_team else None
is_home = player.team is game.state.home_team
for pos in action_choice.positions:
prob = 1
if ball_on_floor_pos is not None:
if pos == ball_on_floor_pos:
prob = 3
elif pos.distance(ball_on_floor_pos) == 1:
prob = 2
elif opp_ball_carrier is not None and pos.distance(
opp_ball_carrier.position):
prob = 2
elif is_ball_carrier and game.arena.is_in_opp_endzone(
pos, is_home):
prob = 2
action = Action(ActionType.MOVE, position=pos)
action_probs.append((action, prob))
return action_probs
def _make_plan(self, game: botbowl.Game, ball_carrier):
#print("1. Stand up marked players")
for player in self.my_team.players:
if player.position is not None and not player.state.up and not player.state.stunned and not player.state.used:
if game.num_tackle_zones_in(player) > 0:
self.actions.append(
Action(ActionType.START_MOVE, player=player))
self.actions.append(Action(ActionType.STAND_UP))
#print(f"Stand up marked player {player.role.name}")
return
#print("2. Move ball carrier to endzone")
if ball_carrier is not None and ball_carrier.team == self.my_team and not ball_carrier.state.used:
#print("2.1 Can ball carrier score with high probability")
td_path = pf.get_safest_path_to_endzone(game,
ball_carrier,
allow_team_reroll=True)
if td_path is not None and td_path.prob >= 0.7:
self.actions.append(
Action(ActionType.START_MOVE, player=ball_carrier))
self.actions.extend(
path_to_move_actions(game, ball_carrier, td_path))
#print(f"Score with ball carrier, p={td_path.prob}")
return
#print("2.2 Hand-off action to scoring player")
if game.is_handoff_available():
# Get players in scoring range
unused_teammates = []
for player in self.my_team.players:
if player.position is not None and player != ball_carrier and not player.state.used and player.state.up:
unused_teammates.append(player)
# Find other players in scoring range
handoff_p = None
handoff_path = None
for player in unused_teammates:
if game.get_distance_to_endzone(
player) > player.num_moves_left():
continue
td_path = pf.get_safest_path_to_endzone(
game, player, allow_team_reroll=True)
if td_path is None:
continue
handoff_path = pf.get_safest_path(game,
ball_carrier,
player.position,
allow_team_reroll=True)
if handoff_path is None:
continue
p_catch = game.get_catch_prob(player,
handoff=True,
allow_catch_reroll=True,
allow_team_reroll=True)
p = td_path.prob * handoff_path.prob * p_catch
if handoff_p is None or p > handoff_p:
handoff_p = p
handoff_path = handoff_path
# Hand-off if high probability or last turn
if handoff_path is not None and (handoff_p >= 0.7 or
self.my_team.state.turn == 8):
self.actions.append(
Action(ActionType.START_HANDOFF, player=ball_carrier))
self.actions.extend(
path_to_move_actions(game, ball_carrier, handoff_path))
return
#print("2.3 Move safely towards the endzone")
if game.num_tackle_zones_in(ball_carrier) == 0:
paths = pf.get_all_paths(game, ball_carrier)
best_path = None
best_distance = 100
target_x = game.get_opp_endzone_x(self.my_team)
for path in paths:
distance_to_endzone = abs(target_x - path.steps[-1].x)
if path.prob == 1 and (
best_path is None or distance_to_endzone <
best_distance) and game.num_tackle_zones_at(
ball_carrier, path.get_last_step()) == 0:
best_path = path
best_distance = distance_to_endzone
if best_path is not None:
self.actions.append(
Action(ActionType.START_MOVE, player=ball_carrier))
self.actions.extend(
path_to_move_actions(game, ball_carrier, best_path))
#print(f"Move ball carrier {ball_carrier.role.name}")
return
#print("3. Safe blocks")
attacker, defender, p_self_up, p_opp_down, block_p_fumble_self, block_p_fumble_opp = self._get_safest_block(
game)
if attacker is not None and p_self_up > 0.94 and block_p_fumble_self == 0:
self.actions.append(Action(ActionType.START_BLOCK,
player=attacker))
self.actions.append(
Action(ActionType.BLOCK, position=defender.position))
#print(f"Safe block with {attacker.role.name} -> {defender.role.name}, p_self_up={p_self_up}, p_opp_down={p_opp_down}")
return
#print("4. Pickup ball")
if game.get_ball_carrier() is None:
pickup_p = None
pickup_player = None
pickup_path = None
for player in self.my_team.players:
if player.position is not None and not player.state.used:
if player.position.distance(
game.get_ball_position()) <= player.get_ma() + 2:
path = pf.get_safest_path(game, player,
game.get_ball_position())
if path is not None:
p = path.prob
if pickup_p is None or p > pickup_p:
pickup_p = p
pickup_player = player
pickup_path = path
if pickup_player is not None and pickup_p > 0.33:
self.actions.append(
Action(ActionType.START_MOVE, player=pickup_player))
self.actions.extend(
path_to_move_actions(game, pickup_player, pickup_path))
#print(f"Pick up the ball with {pickup_player.role.name}, p={pickup_p}")
# Find safest path towards endzone
if game.num_tackle_zones_at(
pickup_player, game.get_ball_position(
)) == 0 and game.get_opp_endzone_x(
self.my_team) != game.get_ball_position().x:
paths = pf.get_all_paths(
game,
pickup_player,
from_position=game.get_ball_position(),
num_moves_used=len(pickup_path))
best_path = None
best_distance = 100
target_x = game.get_opp_endzone_x(self.my_team)
for path in paths:
distance_to_endzone = abs(target_x - path.steps[-1].x)
if path.prob == 1 and (
best_path is None or distance_to_endzone <
best_distance) and game.num_tackle_zones_at(
pickup_player, path.get_last_step()) == 0:
best_path = path
best_distance = distance_to_endzone
if best_path is not None:
self.actions.extend(
path_to_move_actions(game,
pickup_player,
best_path,
do_assertions=False))
#print(f"- Move ball carrier {pickup_player.role.name}")
return
# Scan for unused players that are not marked
open_players = []
for player in self.my_team.players:
if player.position is not None and not player.state.used and game.num_tackle_zones_in(
player) == 0:
open_players.append(player)
#print("5. Move receivers into scoring distance if not already")
for player in open_players:
if player.has_skill(Skill.CATCH) and player != ball_carrier:
if game.get_distance_to_endzone(
player) > player.num_moves_left():
continue
paths = pf.get_all_paths(game, player)
best_path = None
best_distance = 100
target_x = game.get_opp_endzone_x(self.my_team)
for path in paths:
distance_to_endzone = abs(target_x - path.steps[-1].x)
if path.prob == 1 and (best_path is None or
distance_to_endzone < best_distance
) and game.num_tackle_zones_at(
player, path.get_last_step()):
best_path = path
best_distance = distance_to_endzone
if best_path is not None:
self.actions.append(
Action(ActionType.START_MOVE, player=player))
self.actions.extend(
path_to_move_actions(game, player, best_path))
#print(f"Move receiver {player.role.name}")
return
#print("6. Blitz with open block players")
if game.is_blitz_available():
best_blitz_attacker = None
best_blitz_defender = None
best_blitz_score = None
best_blitz_path = None
for blitzer in open_players:
if blitzer.position is not None and not blitzer.state.used and blitzer.has_skill(
Skill.BLOCK):
blitz_paths = pf.get_all_paths(game, blitzer, blitz=True)
for path in blitz_paths:
defender = game.get_player_at(path.get_last_step())
if defender is None:
continue
from_position = path.steps[-2] if len(
path.steps) > 1 else blitzer.position
p_self, p_opp, p_fumble_self, p_fumble_opp = game.get_blitz_probs(
blitzer, from_position, defender)
p_self_up = path.prob * (1 - p_self)
p_opp = path.prob * p_opp
p_fumble_opp = p_fumble_opp * path.prob
if blitzer == game.get_ball_carrier():
p_fumble_self = path.prob + (
1 - path.prob) * p_fumble_self
score = p_self_up + p_opp + p_fumble_opp - p_fumble_self
if best_blitz_score is None or score > best_blitz_score:
best_blitz_attacker = blitzer
best_blitz_defender = defender
best_blitz_score = score
best_blitz_path = path
if best_blitz_attacker is not None and best_blitz_score >= 1.25:
self.actions.append(
Action(ActionType.START_BLITZ, player=best_blitz_attacker))
self.actions.extend(
path_to_move_actions(game, best_blitz_attacker,
best_blitz_path))
#print(f"Blitz with {best_blitz_attacker.role.name}, score={best_blitz_score}")
return
#print("7. Make cage around ball carrier")
cage_positions = [
Square(game.get_ball_position().x - 1,
game.get_ball_position().y - 1),
Square(game.get_ball_position().x + 1,
game.get_ball_position().y - 1),
Square(game.get_ball_position().x - 1,
game.get_ball_position().y + 1),
Square(game.get_ball_position().x + 1,
game.get_ball_position().y + 1)
]
if ball_carrier is not None:
for cage_position in cage_positions:
if game.get_player_at(
cage_position) is None and not game.is_out_of_bounds(
cage_position):
for player in open_players:
if player == ball_carrier or player.position in cage_positions:
continue
if player.position.distance(
cage_position) > player.num_moves_left():
continue
if game.num_tackle_zones_in(player) > 0:
continue
path = pf.get_safest_path(game, player, cage_position)
if path is not None and path.prob > 0.94:
self.actions.append(
Action(ActionType.START_MOVE, player=player))
self.actions.extend(
path_to_move_actions(game, player, path))
#print(f"Make cage around towards ball carrier {player.role.name}")
return
# Scan for assist positions
assist_positions = set()
for player in game.get_opp_team(self.my_team).players:
if player.position is None or not player.state.up:
continue
for opponent in game.get_adjacent_opponents(player, down=False):
att_str, def_str = game.get_block_strengths(player, opponent)
if def_str >= att_str:
for open_position in game.get_adjacent_squares(
player.position, occupied=False):
if len(
game.get_adjacent_players(open_position,
team=self.opp_team,
down=False)) == 1:
assist_positions.add(open_position)
#print("8. Move non-marked players to assist")
for player in open_players:
for path in pf.get_all_paths(game, player):
if path.prob < 1.0 or path.get_last_step(
) not in assist_positions:
continue
self.actions.append(
Action(ActionType.START_MOVE, player=player))
self.actions.extend(path_to_move_actions(game, player, path))
# print(f"Move assister {player.role.name} to {path.get_last_step().to_json}")
return
#print("9. Move towards the ball")
for player in open_players:
if player == ball_carrier or game.num_tackle_zones_in(player) > 0:
continue
shortest_distance = None
path = None
if ball_carrier is None:
for p in pf.get_all_paths(game, player):
distance = p.get_last_step().distance(
game.get_ball_position())
if shortest_distance is None or (
p.prob == 1 and distance < shortest_distance):
shortest_distance = distance
path = p
elif ball_carrier.team != self.my_team:
for p in pf.get_all_paths(game, player):
distance = p.get_last_step().distance(
ball_carrier.position)
if shortest_distance is None or (
p.prob == 1 and distance < shortest_distance):
shortest_distance = distance
path = p
if path is not None:
self.actions.append(
Action(ActionType.START_MOVE, player=player))
self.actions.extend(path_to_move_actions(game, player, path))
#print(f"Move towards ball {player.role.name}")
return
#print("10. Risky blocks")
attacker, defender, p_self_up, p_opp_down, block_p_fumble_self, block_p_fumble_opp = self._get_safest_block(
game)
if attacker is not None and (p_opp_down > (1 - p_self_up)
or block_p_fumble_opp > 0):
self.actions.append(Action(ActionType.START_BLOCK,
player=attacker))
self.actions.append(
Action(ActionType.BLOCK, position=defender.position))
#print(f"Block with {player.role.name} -> {defender.role.name}, p_self_up={p_self_up}, p_opp_down={p_opp_down}")
return
#print("11. End turn")
self.actions.append(Action(ActionType.END_TURN))
def hash_game_state(game: botbowl.Game) -> int:
s = "".join([f"{p.position.x}{p.position.y}{int(p.state.up)}"
for p in game.get_players_on_pitch()])
s += "".join([f"{ball_pos.x}{ball_pos.y}" for ball_pos in game.get_ball_positions()])
return hash(s)
def expand_block(game: botbowl.Game, parent: Node) -> Node:
proc: botbowl.Block = game.get_procedure()
assert type(proc) is botbowl.Block
assert not proc.gfi, "Can't handle GFI:s here =( "
assert proc.roll is None
attacker: botbowl.Player = proc.attacker
defender: botbowl.Player = proc.defender
dice = game.num_block_dice(attacker, defender)
num_dice = abs(dice)
# initialize as 1d block without skills
dice_outcomes = np.array([2, 2, 1, 1], dtype=int)
DEF_DOWN, NOONE_DOWN, ALL_DOWN, ATT_DOWN = (0, 1, 2, 3)
die_results = ([BBDieResult.DEFENDER_DOWN,
BBDieResult.DEFENDER_STUMBLES], [BBDieResult.PUSH],
[BBDieResult.BOTH_DOWN], [BBDieResult.ATTACKER_DOWN])
who_has_block = (attacker.has_skill(Skill.BLOCK),
defender.has_skill(Skill.BLOCK))
if any(who_has_block):
dice_outcomes[ALL_DOWN] = 0
die_results[ALL_DOWN].clear()
if who_has_block == (True, True): # both
dice_outcomes[NOONE_DOWN] += 1
die_results[NOONE_DOWN].append(BBDieResult.BOTH_DOWN)
elif who_has_block == (True, False): # only attacker
dice_outcomes[DEF_DOWN] += 1
die_results[DEF_DOWN].append(BBDieResult.BOTH_DOWN)
elif who_has_block == (False, True): # only defender
dice_outcomes[ATT_DOWN] += 1
die_results[ATT_DOWN].append(BBDieResult.BOTH_DOWN)
crowd_surf: bool = game.get_push_squares(
attacker.position, defender.position)[0].out_of_bounds
if crowd_surf:
dice_outcomes[DEF_DOWN] += 2
dice_outcomes[NOONE_DOWN] -= 2
die_results[DEF_DOWN].append(BBDieResult.PUSH)
die_results[NOONE_DOWN].remove(BBDieResult.PUSH)
elif defender.has_skill(
Skill.DODGE): # and not attacker.has_skill(Skill.TACKLE):
dice_outcomes[DEF_DOWN] -= 1
dice_outcomes[NOONE_DOWN] += 1
die_results[DEF_DOWN].remove(BBDieResult.DEFENDER_STUMBLES)
die_results[NOONE_DOWN].append(BBDieResult.DEFENDER_STUMBLES)
prob = np.zeros(4)
probability_left = 1.0
available_dice = 6
evaluation_order = [DEF_DOWN, NOONE_DOWN, ALL_DOWN, ATT_DOWN]
if dice < 0:
evaluation_order = reversed(evaluation_order)
for i in evaluation_order:
prob[i] = probability_left * (
1 - (1 - dice_outcomes[i] / available_dice)**num_dice)
available_dice -= dice_outcomes[i]
probability_left -= prob[i]
assert available_dice == 0 and probability_left == approx(
0) and prob.sum() == approx(1)
new_parent = ChanceNode(game, parent)
for prob, die_res in zip(prob, die_results):
if prob == approx(0) or len(die_res) == 0:
assert prob == approx(0) and len(die_res) == 0
continue
expand_with_fixes(game,
new_parent,
prob,
block_dice=np.random.choice(die_res, num_dice))
assert sum(new_parent.child_probability) == approx(1.0)
return new_parent
def expand_moving(game: botbowl.Game, parent: Node) -> Node:
# noinspection PyTypeChecker
active_proc: Union[procedures.GFI, procedures.Dodge] = game.get_procedure()
assert type(active_proc) is procedures.Dodge or type(
active_proc) is procedures.GFI
move_action_proc: procedures.MoveAction = first(
proc for proc in reversed(game.state.stack.items)
if isinstance(proc, procedures.MoveAction))
is_blitz = type(move_action_proc) is procedures.BlitzAction
is_handoff = type(move_action_proc) is procedures.HandoffAction
player = move_action_proc.player
if move_action_proc.steps is not None:
final_step = move_action_proc.steps[-1]
else:
if is_blitz:
block_proc: procedures.Block = first(
filter(lambda proc: type(proc) is procedures.Block,
game.state.stack.items))
final_step = block_proc.defender.position
elif is_handoff:
raise ValueError()
else:
final_step = active_proc.position
is_pickup = game.get_ball(
).position == final_step and not game.get_ball().is_carried
path = move_action_proc.paths[final_step]
"""
This block of code sets two important variables:
probability_success - probability of the remaining path
rolls - list[int] - the remaining rolls of the path
Normal case we just fetch this from the path object. If we're in a rerolled proc, it's nasty...
"""
if active_proc.roll is None:
probability_success = path.prob
rolls = list(collapse(path.rolls))
if is_pickup:
# remove the pickup roll and probability
rolls.pop()
probability_success /= game.get_pickup_prob(
active_proc.player, final_step)
else:
with only_fixed_rolls(game):
game.step()
new_proc = game.get_procedure()
if type(new_proc) not in {procedures.GFI, procedures.Dodge}:
assert not active_proc.reroll.use_reroll
return expand_none_action(game, parent)
# if we get here, it means that a reroll was used.
assert new_proc is active_proc
assert active_proc.roll is None
assert active_proc.reroll is None
current_step = active_proc.position
try:
assert player.position.distance(
current_step) == 1 or is_pickup or is_blitz
except AssertionError as e:
raise e
i = 0
while path.steps[i] != current_step:
i += 1
remaining_current_step_rolls = path.rolls[i][:]
if is_pickup and current_step == final_step:
remaining_current_step_rolls.pop()
num_current_step_remaining_rolls = len(
list(
filter(lambda x: type(x) in {procedures.GFI, procedures.Dodge},
game.state.stack.items)))
assert num_current_step_remaining_rolls in [1, 2]
remaining_current_step_rolls = remaining_current_step_rolls[
-num_current_step_remaining_rolls:]
probability_success = reduce(
operator.mul,
map(lambda d: (7 - d) / 6, remaining_current_step_rolls), 1.0)
rolls = list(collapse(remaining_current_step_rolls))
if current_step != final_step:
step_count = game.get_step()
if player.position != current_step:
try:
game.move(player, current_step)
except AssertionError as e:
raise e
new_path = pf.get_safest_path(game,
player,
final_step,
blitz=is_blitz)
game.revert(step_count)
#try:
# # assert new_path.steps == path.steps[-len(new_path):] this assert can't be made because of small randomness in pathfinder
# assert list(collapse(new_path.rolls)) == list(collapse(path.rolls[-len(new_path):])), f"{new_path.rolls} != {path.rolls[-len(new_path):]}"
#except AssertionError as e:
# raise e
try:
rolls.extend(collapse(new_path.rolls))
except AttributeError as e:
raise e
probability_success *= new_path.prob
if is_pickup:
# remove the pickup roll and probability
rolls.pop()
probability_success /= game.get_pickup_prob(
active_proc.player, final_step)
try:
p = np.array(rolls) / sum(rolls)
index_of_failure = np.random.choice(range(len(rolls)), 1, p=p)[0]
except ValueError as e:
raise e
# STEP UNTIL FAILURE (possibly no steps at all)
with only_fixed_rolls(game, d6=[6] * index_of_failure):
while len(botbowl.D6.FixedRolls) > 0:
if len(game.get_available_actions()) > 0:
raise AttributeError("wrong")
game.step()
new_parent = ChanceNode(game, parent)
debug_step_count = game.get_step()
# SUCCESS SCENARIO
with only_fixed_rolls(game, d6=[6] * (len(rolls) - index_of_failure)):
while len(botbowl.D6.FixedRolls) > 0:
if type(game.get_procedure()) not in {
procedures.GFI, procedures.Block, procedures.Dodge,
procedures.Move, procedures.MoveAction,
procedures.BlitzAction
}:
raise AttributeError("wrong")
if len(game.get_available_actions()) > 0:
raise AttributeError("wrong")
if type(game.get_procedure()
) is procedures.Block and not game.get_procedure().gfi:
raise AttributeError("wrong")
game.step()
success_node = expand_none_action(game, new_parent, moving_handled=True)
new_parent.connect_child(success_node, probability_success)
assert debug_step_count == game.get_step()
# FAILURE SCENARIO
with only_fixed_rolls(game, d6=[1]):
while len(botbowl.D6.FixedRolls) > 0:
if len(game.get_available_actions()) > 0:
raise AttributeError("wrong")
game.step()
fail_node = expand_none_action(game, new_parent, moving_handled=True)
new_parent.connect_child(fail_node, 1 - probability_success)
assert debug_step_count == game.get_step()
return new_parent