next_pos[ship.id].x][2]]
else:
continue
# print(f'Stopped ship id {ship.id} to prevent collision')
next_ships[next_pos[ship.id].y][next_pos[ship.id].x].remove(
ship)
next_pos[ship.id].x = ship.x
next_pos[ship.id].y = ship.y
commands[ship.id] = MoveCommand(self.id, ship.id, 'O')
q.extend(next_ships[ship.y][ship.x])
next_ships[ship.y][ship.x].append(ship)
ret = list(commands.values())
if (len(next_ships[self.shipyard.y][self.shipyard.x]) == 0
and self.halite >= (1000 if self.pd is None else 5000)
and game.max_turns - game.turn > 100):
ret.append(SpawnShipCommand(self.id, None))
return ret
if __name__ == '__main__':
bot1 = FastBot()
bot2 = StandardBot()
players, cell_data, bank_data, owner_data, collisions = Game.run_game(
[bot1, bot2], return_replay=True, map_gen='perlin')
my_replay = Replayer.from_data(players, cell_data, bank_data, owner_data,
collisions)
my_replay.run()
ship.id].x].remove(ship)
next_pos[ship.id].x = ship.x
next_pos[ship.id].y = ship.y
commands[ship.id] = MoveCommand(self.id, ship.id, 'O')
q.extend(next_ships[ship.y][ship.x])
next_ships[ship.y][ship.x].append(ship)
ret = list(commands.values())
if (len(next_ships[self.shipyard.y][self.shipyard.x]) == 0
and self.halite >= 1000 and rem > 100 and
np.sum(game.cells[:, :, 0]) * 3 > self.map_starting_halite):
ret.append(SpawnShipCommand(self.id, None))
return ret
def __repr__(self):
return f'{self.__class__.__name__}(id={self.id}, eps={self.eps}, training={self.memory is not None})'
epsilon = MAX_EPSILON
model = create_unet()
model.load_weights('./checkpoints/cp5000_02.ckpt')
mem = Memory(10000)
for step_num in range(500):
epsilon = MIN_EPSILON + (MAX_EPSILON - MIN_EPSILON) * math.exp(
-LAMBDA * step_num)
rl_bot = RLBot(model, memory=mem, eps=epsilon)
Game.run_game([rl_bot, FastBot()], map_width=32, verbosity=0)
print(f'Loss: {rl_bot.total_loss}')
if step_num % 20 == 19:
model.save_weights(f'./checkpoints/RL_checkpoint_{step_num}')
NUM_TRAIN = 1000 # number of full validation games
SPARSE_FREQ = 0.02
DENSE_FREQ = 0.02
EPS = 0.05
NUM_VAL = 50
TRAIN_RATIO = 0.2 # proportion of training turns saved (number of games ran adjusted so total turns saved is same)
VAL_RATIO = 0.1 # proportion of validation turns saved
for i in range(3100, int(NUM_TRAIN / TRAIN_RATIO)):
x = random.random()
if x < SPARSE_FREQ:
print(
f'Creating sparse training game {i + 1:03}/{int(NUM_TRAIN / TRAIN_RATIO)}'
)
Game.run_game([FastBot(eps=EPS), FastBot(eps=EPS)],
map_width=32,
map_gen='sparse',
save_name=f'data/game{i}',
save_split=TRAIN_RATIO,
verbosity=0)
elif x < SPARSE_FREQ + DENSE_FREQ:
print(
f'Creating dense training game {i + 1:03}/{int(NUM_TRAIN / TRAIN_RATIO)}'
)
Game.run_game([FastBot(eps=EPS), FastBot(eps=EPS)],
map_width=32,
map_gen='dense',
save_name=f'data/game{i}',
save_split=TRAIN_RATIO,
verbosity=0)
else:
print(
f'Creating perlin training game {i+1:03}/{int(NUM_TRAIN / TRAIN_RATIO)}'