def create_experiment_3(overwrite: bool = True, show: bool = True):
"""
Experiment 3 tests both the generalization capabilities in adapting to newly spawned targets. In this scenario, new
targets are spawned ones the previous are found, with the goal to find as many as possible.
TRAINING
During training, new targets are spawned based on the agent's current position. Each target will lay between 4 and
8 meters from the agent and will always be positioned inside of the maze.
TESTING
Testing is done similar as was the case for training, with the main difference that the positions of the targets
are defined beforehand, this to ensure a fair comparison between the different evaluations. 10 targets are created,
with the idea that it would be impossible for the drone to ever reach this many targets.
"""
cfg = get_shared_config()
# Fixed parameters
ROBOT_INIT_ANGLE = pi / 2 # Looking up
ROBOT_INIT_POS = Vec2d(cfg.game.x_axis / 2, cfg.game.y_axis /
2) # Initial position of the drone
# Create the training game
spawn_f_train = SpawnRandom(game_config=cfg.game, train=True)
game = Game(
config=cfg,
player_noise=0,
game_id=get_game_id(0, experiment_id=3),
overwrite=overwrite,
spawn_func=spawn_f_train,
stop_if_reached=False,
wall_bound=False,
)
game.player = Robot(game=game)
game.set_player_init_angle(a=ROBOT_INIT_ANGLE)
game.set_player_init_pos(p=ROBOT_INIT_POS)
check_and_save_game(game, show=show)
# Create 20 evaluation games
for i in range(1, 21):
spawn_f_eval = SpawnRandom(game_config=cfg.game, train=False)
game = Game(
config=cfg,
player_noise=0,
game_id=get_game_id(i, experiment_id=3),
overwrite=overwrite,
spawn_func=spawn_f_eval,
stop_if_reached=False,
wall_bound=False,
)
game.player = Robot(game=game)
game.set_player_init_angle(a=ROBOT_INIT_ANGLE)
game.set_player_init_pos(p=ROBOT_INIT_POS)
check_and_save_game(game, show=show, randomize=False)
def create_dummy_game(i: int = 0, overwrite: bool = True, show: bool = True):
cfg = get_shared_config()
ROBOT_INIT_POS = Vec2d(cfg.game.x_axis / 2, cfg.game.y_axis /
2) # Initial position of the drone
spawn_f = SpawnSimple(game_config=cfg, train=True)
spawn_f.add_location((1, 1))
game = Game(
config=cfg,
game_id=get_game_id(i, experiment_id=0),
overwrite=overwrite,
spawn_func=spawn_f,
)
game.player = Robot(game=game)
game.set_player_init_pos(p=ROBOT_INIT_POS)
check_and_save_game(game, show=show)
def get_game():
cfg = Config()
cfg.game.fps = 60
cfg.game.x_axis = 1
cfg.game.y_axis = 1
cfg.update()
spawn = SpawnSimple(game_config=cfg)
spawn.add_location((10, 10)) # Dummy location
game = Game(
config=cfg,
game_id=0,
noise=False,
overwrite=True,
save_path="tests/games_db/",
silent=True,
spawn_func=spawn,
wall_bound=True,
)
game.set_player_init_pos(Vec2d(0.5, 0.5))
game.sample_target()
game.reset()
return game
def create_experiment_1(overwrite: bool = True, show: bool = True):
"""
Experiment 1 tests the generalization capabilities of the genomes based on the relative direction the target is
placed.
TRAINING
Training is done in a maze where the robot is in the center, and the targets are positioned around the agent. Each
run has a single target sampled. The targets all are positioned 6 meters from the robot's initial position.
TESTING
Testing is done similar as was the case for training, however, in this scenario only 20 targets are sampled (where
there were 36 for training). Only 4 of these 20 sampled target positions overlap with the training set.
"""
cfg = get_shared_config()
# Fixed parameters
ROBOT_INIT_ANGLE = pi / 2 # Looking up
ROBOT_INIT_POS = Vec2d(cfg.game.x_axis / 2, cfg.game.y_axis /
2) # Initial position of the drone
# Create the training game
spawn_f_train = SpawnSimple(game_config=cfg, train=True)
for i in range(0, 360, 10): # Positions circular in hops of 10 degree
angle = i / 180 * pi
offset_x = 6 * cos(angle)
offset_y = 6 * sin(angle)
spawn_f_train.add_location(
(ROBOT_INIT_POS[0] + offset_x, ROBOT_INIT_POS[1] + offset_y))
game = Game(
config=cfg,
player_noise=0,
game_id=get_game_id(0, experiment_id=1),
overwrite=overwrite,
spawn_func=spawn_f_train,
stop_if_reached=True,
wall_bound=False,
)
game.player = Robot(game=game)
game.set_player_init_angle(a=ROBOT_INIT_ANGLE)
game.set_player_init_pos(p=ROBOT_INIT_POS)
check_and_save_game(game, show=show)
# Create the evaluation games, each of those only contains one target
for i in range(1, 21):
spawn_f_eval = SpawnSimple(game_config=cfg, train=False)
angle = i / 10 * pi
offset_x = 6 * cos(angle)
offset_y = 6 * sin(angle)
spawn_f_eval.add_location(
(ROBOT_INIT_POS[0] + offset_x, ROBOT_INIT_POS[1] + offset_y))
game = Game(
config=cfg,
player_noise=0,
game_id=get_game_id(i, experiment_id=1),
overwrite=overwrite,
spawn_func=spawn_f_eval,
stop_if_reached=True,
wall_bound=False,
)
game.player = Robot(game=game)
game.set_player_init_angle(a=ROBOT_INIT_ANGLE)
game.set_player_init_pos(p=ROBOT_INIT_POS)
check_and_save_game(game, show=show, randomize=False)
def create_experiment_6(overwrite: bool = True, show: bool = True):
"""
Experiment 6 is another 'simplified' simulation in which the agents are trained on a harder environment than they
are tested on. The reason why is to keep a lower threshold for the 'solution' genomes (only interested in those).
TRAINING
During training, new targets are spawned based on the agent's initial position. Each target will lay between 2 and
10 meters from the agent and will always be positioned inside of the maze.
TESTING
Testing is done on 18 predefined targets position in a relative angle from the agent of k*20° and a distance of
either 4, 6, or 8 meters (sequential alternating).
"""
cfg = get_shared_config()
# Fixed parameters
ROBOT_INIT_ANGLE = pi / 2 # Looking up
ROBOT_INIT_POS = Vec2d(cfg.game.x_axis / 2, cfg.game.y_axis /
2) # Initial position of the drone
# Create the training game
spawn_f_train = SpawnRange(game_config=cfg.game,
train=True,
r_max=8,
r_min=4)
game = Game(
config=cfg,
player_noise=0,
game_id=get_game_id(0, experiment_id=6),
overwrite=overwrite,
spawn_func=spawn_f_train,
stop_if_reached=True,
wall_bound=False,
)
game.player = Robot(game=game)
game.set_player_init_angle(a=ROBOT_INIT_ANGLE)
game.set_player_init_pos(p=ROBOT_INIT_POS)
check_and_save_game(game, show=show)
# Create 18 evaluation games
for i in range(1, 19):
spawn_f_eval = SpawnSimple(game_config=cfg.game, train=False)
angle = i / 9 * pi # Hops of 20°
d = 4 if i % 3 == 0 else 6 if i % 3 == 1 else 8
offset_x = d * cos(angle)
offset_y = d * sin(angle)
spawn_f_eval.add_location(
(ROBOT_INIT_POS[0] + offset_x, ROBOT_INIT_POS[1] + offset_y))
game = Game(
config=cfg,
player_noise=0,
game_id=get_game_id(i, experiment_id=6),
overwrite=overwrite,
spawn_func=spawn_f_eval,
stop_if_reached=True,
wall_bound=False,
)
game.player = Robot(game=game)
game.set_player_init_angle(a=ROBOT_INIT_ANGLE)
game.set_player_init_pos(p=ROBOT_INIT_POS)
check_and_save_game(game, show=show, randomize=False)
def create_experiment_2(overwrite: bool = True, show: bool = True):
"""
Experiment 2 tests the generalization capabilities of the genomes based on their initial distance from the target.
TRAINING
Training is done in a maze where the robot is in the center, and the targets are positioned around the agent. Each
run has a single target sampled. The targets all are positioned 6 meters from the robot's initial position.
TESTING
Testing is done similar as was the case for training. For testing, half of the training's initial target-
orientations are used (18 instead of 36). However, each orientation now has two targets, both with a different
initial position from the robot's starting position. An inner-ring has distance 4, where the outer-ring has a
distance of 8.
"""
cfg = get_shared_config()
# Fixed parameters
ROBOT_INIT_ANGLE = pi / 2 # Looking up
ROBOT_INIT_POS = Vec2d(cfg.game.x_axis / 2, cfg.game.y_axis /
2) # Initial position of the drone
# Create the training game
spawn_f_train = SpawnSimple(game_config=cfg, train=True)
for i in range(0, 360, 10): # Positions circular in hops of 10 degree
angle = i / 180 * pi
offset_x = 6 * cos(angle)
offset_y = 6 * sin(angle)
spawn_f_train.add_location(
(ROBOT_INIT_POS[0] + offset_x, ROBOT_INIT_POS[1] + offset_y))
game = Game(
config=cfg,
player_noise=0,
game_id=get_game_id(0, experiment_id=2),
overwrite=overwrite,
spawn_func=spawn_f_train,
stop_if_reached=True,
wall_bound=False,
)
game.player = Robot(game=game)
game.set_player_init_angle(a=ROBOT_INIT_ANGLE)
game.set_player_init_pos(p=ROBOT_INIT_POS)
check_and_save_game(game, show=show)
# Create the inner circle of the evaluation games, each of those only contains one target
for i in range(1, 19):
spawn_f_eval = SpawnSimple(game_config=cfg, train=False)
angle = i / 9 * pi # Hops of 20°
offset_x = 4 * cos(angle)
offset_y = 4 * sin(angle)
spawn_f_eval.add_location(
(ROBOT_INIT_POS[0] + offset_x, ROBOT_INIT_POS[1] + offset_y))
game = Game(
config=cfg,
player_noise=0,
game_id=get_game_id(i, experiment_id=2),
overwrite=overwrite,
spawn_func=spawn_f_eval,
stop_if_reached=True,
wall_bound=False,
)
game.player = Robot(game=game)
game.set_player_init_angle(a=ROBOT_INIT_ANGLE)
game.set_player_init_pos(p=ROBOT_INIT_POS)
check_and_save_game(game, show=show, randomize=False)
# Create the outer circle of the evaluation games, each of those only contains one target
for i in range(1, 19):
spawn_f_eval = SpawnSimple(game_config=cfg, train=False)
angle = i / 9 * pi # Hops of 20°
offset_x = 8 * cos(angle)
offset_y = 8 * sin(angle)
spawn_f_eval.add_location(
(ROBOT_INIT_POS[0] + offset_x, ROBOT_INIT_POS[1] + offset_y))
game = Game(
config=cfg,
player_noise=0,
game_id=get_game_id(i + 100, experiment_id=2),
overwrite=overwrite,
spawn_func=spawn_f_eval,
stop_if_reached=True,
wall_bound=False,
)
game.player = Robot(game=game)
game.set_player_init_angle(a=ROBOT_INIT_ANGLE)
game.set_player_init_pos(p=ROBOT_INIT_POS)
check_and_save_game(game, show=show, randomize=False)