def __init__(self, name, team, index):
super().__init__(name, team, index)
import torch
sys.path.insert(0, os.path.dirname(
os.path.abspath(__file__))) # this is for separate process imports
from examples.levi.torch_model import SymmetricModel
self.ga = GeneticAlgorithm()
self.Model = SymmetricModel
self.torch = torch
self.controller_state = SimpleControllerState()
self.frame = 0 # frame counter for timed reset
self.brain = 0 # bot counter for generation reset
self.pop = 10 # population for bot looping
self.num_best = 5
self.gen = 0
self.pos = 0
self.max_frames = 5000
self.bot_list = [self.Model() for _ in range(self.pop)
] # list of Individual() objects
self.bot_list[-self.num_best:] = [
self.Model()
] * self.num_best # make sure last bots are the same
self.bot_fitness = [0] * self.pop
self.fittest = None # fittest object
self.mut_rate = 0.2 # mutation rate
self.distance_to_ball = [
math.inf
] * self.max_frames # set high for easy minimum
self.input_formatter = LeviInputFormatter(team, index)
self.output_formatter = LeviOutputFormatter(index)
class CoolAtbaAgent(BaseAgent):
input_formatter = None
atba = None
output_formatter = None
def __init__(self, name, team, index):
super().__init__(name, team, index)
self.logger = get_logger(name)
def initialize_agent(self):
self.input_formatter = LeviInputFormatter(self.team, self.index)
self.atba = Atba()
self.output_formatter = LeviOutputFormatter(self.index)
def get_output(self, game_tick_packet: GameTickPacket):
arr = self.input_formatter.create_input_array([game_tick_packet])
output = self.atba.get_action(arr)
return self.output_formatter.format_model_output(
output, game_tick_packet)[0]
def initialize_agent(self):
self.input_formatter = LeviInputFormatter(self.team, self.index)
self.atba = Atba()
self.output_formatter = LeviOutputFormatter(self.index)
def create_input_formatter(self):
return LeviInputFormatter(self.team, self.index)
class SelfEvolvingCar(BaseAgent):
"""This agent uses neuro-evolution to train the Levi torch model to perform aerials in Rocket League.
first, the algorithm runs each model with randomly generated parameters. It then calculates each bots' fitness
by determining it's minimum distance to the ball. Next, it clones the bot with the best fitness to the rest of
the network, and uses a mutation function to guarantee diversity in the population. Make sure to change match
length and points to unlimited and disable goal reset and enable instant start"""
def __init__(self, name, team, index):
super().__init__(name, team, index)
import torch
sys.path.insert(0, os.path.dirname(
os.path.abspath(__file__))) # this is for separate process imports
from examples.levi.torch_model import SymmetricModel
self.ga = GeneticAlgorithm()
self.Model = SymmetricModel
self.torch = torch
self.controller_state = SimpleControllerState()
self.frame = 0 # frame counter for timed reset
self.brain = 0 # bot counter for generation reset
self.pop = 10 # population for bot looping
self.num_best = 5
self.gen = 0
self.pos = 0
self.max_frames = 5000
self.bot_list = [self.Model() for _ in range(self.pop)
] # list of Individual() objects
self.bot_list[-self.num_best:] = [
self.Model()
] * self.num_best # make sure last bots are the same
self.bot_fitness = [0] * self.pop
self.fittest = None # fittest object
self.mut_rate = 0.2 # mutation rate
self.distance_to_ball = [
math.inf
] * self.max_frames # set high for easy minimum
self.input_formatter = LeviInputFormatter(team, index)
self.output_formatter = LeviOutputFormatter(index)
def initialize_agent(self):
self.reset() # reset at start
def get_output(self, packet: GameTickPacket) -> SimpleControllerState:
# neural net inputs
inputs = self.input_formatter.create_input_array([packet])
inputs = [self.torch.from_numpy(x).float() for x in inputs]
# fitness function
my_car = packet.game_cars[self.index]
distance_to_ball_x = packet.game_ball.physics.location.x - my_car.physics.location.x
distance_to_ball_y = packet.game_ball.physics.location.y - my_car.physics.location.y
distance_to_ball_z = packet.game_ball.physics.location.z - my_car.physics.location.z
self.distance_to_ball[self.frame] = math.sqrt(distance_to_ball_x**2 +
distance_to_ball_y**2 +
distance_to_ball_z**2)
# render results
action_display = "GEN: " + str(self.gen + 1) + " | BOT: " + str(
self.brain)
draw_debug(self.renderer, action_display)
# stop evolving when ball is touched
# if packet.game_ball.latest_touch.player_name == "Self-Evolving-Car":
# self.mut_rate = 0
# game state
car_state = CarState(boost_amount=100)
ball_state = BallState(
Physics(velocity=Vector3(0, 0, 0),
location=Vector3(0, -1000, 1200),
angular_velocity=Vector3(0, 0, 0)))
game_state = GameState(ball=ball_state, cars={self.index: car_state})
self.set_game_state(game_state)
# neural net outputs
with self.torch.no_grad():
outputs = self.bot_list[self.brain].forward(*inputs)
self.controller_state = self.output_formatter.format_model_output(
outputs, [packet])[0]
# kill
if (my_car.physics.location.z < 100 or my_car.physics.location.z > 1950
or my_car.physics.location.x < -4000 or my_car.physics.location.x > 4000
or my_car.physics.location.y > 5000 or my_car.physics.location.y < -5000) \
and self.frame > 50:
self.frame = self.max_frames
# loops
self.frame = self.frame + 1
if self.frame >= self.max_frames:
self.frame = 0
self.bot_fitness[self.brain] = self.ga.calc_fitness(
self.distance_to_ball)
self.distance_to_ball = [math.inf] * self.max_frames
self.brain += 1 # change bot every reset
self.controller_state = SimpleControllerState() # reset controller
self.reset() # reset at start
if self.brain >= self.pop:
self.gen += 1
self.brain = 0 # reset bots after all have gone
self.bot_fitness[self.num_best:] = [
self.ga.avg_best_fitness(self.bot_fitness[self.num_best:])
for _ in self.bot_fitness[self.num_best:]
]
self.fittest = self.ga.calc_fittest(self.bot_fitness)
# print generation info
self.logger.info("")
self.logger.info(" GEN = {}".format(self.gen))
self.logger.info("-------------------------")
self.logger.info("FITTEST = BOT {}".format(self.fittest))
self.logger.info("------FITNESS = {}".format(
self.bot_fitness[self.fittest]))
for i in range(len(self.bot_list)):
self.logger.info("FITNESS OF BOT {} = {}".format(
i, self.bot_fitness[i]))
# NE functions
self.ga.crossover(self.bot_list[self.fittest], self.bot_list)
self.ga.mutate(self.bot_list[:self.num_best], self.mut_rate)
self.reset() # reset because of delay
return self.controller_state
def reset(self):
"""Resets game data after each genome"""
ball_state = BallState(
Physics(velocity=Vector3(0, 0, 0),
location=Vector3(self.pos, 5000, 3000),
angular_velocity=Vector3(0, 0, 0)))
car_state = CarState(jumped=False,
double_jumped=False,
boost_amount=33,
physics=Physics(velocity=Vector3(0, 0, 0),
rotation=Rotator(45, 90, 0),
location=Vector3(0.0, -4608, 500),
angular_velocity=Vector3(0, 0,
0)))
game_info_state = GameInfoState(game_speed=1)
game_state = GameState(ball=ball_state,
cars={self.index: car_state},
game_info=game_info_state)
self.set_game_state(game_state)