def get_state():
"""Get initial state."""
sprite_0 = sprite.Sprite(x=0.5,
y=0.7,
scale=0.1,
shape='triangle',
x_vel=0.04,
y_vel=-0.02,
c0=255,
angle=2.)
sprite_1 = sprite.Sprite(x=0.2,
y=0.6,
scale=0.1,
shape='triangle',
x_vel=0.,
y_vel=0.,
c1=255,
angle=1.)
sprite_2 = sprite.Sprite(x=0.6,
y=0.3,
scale=0.1,
shape='triangle',
x_vel=0.,
y_vel=0.,
c2=255)
sprites = [sprite_0, sprite_1, sprite_2]
walls = shapes.border_walls(visible_thickness=0.05, c0=128, c1=128, c2=128)
state = collections.OrderedDict([('walls', walls), ('sprites', sprites)])
return state
def __init__(self, num_predators, step_scaling_factor,
threshold_trial_len):
"""Constructor.
This class uses the meta-state to keep track of the number of steps
before the agent is caught. See the game rules section near the bottom
of this file for the counter incrementer.
Args:
step_scaling_factor: Float. Fractional decrease of predator mass
after a trial longer than threshold_trial_len. Also used as
fractional increase of predator mass after a trial shorter than
threshold_trial_len. Should be small and positive.
threshold_trial_len: Length of a trial above which the predator
mass is decreased and below which the predator mass is
increased.
"""
self._mass = 1.
self._step_scaling_factor = step_scaling_factor
self._threshold_trial_len = threshold_trial_len
# Agent
agent_factors = distribs.Product(
[
distribs.Continuous('x', 0., 1.),
distribs.Continuous('y', 0., 1.)
],
shape='circle',
scale=0.1,
c0=0.33,
c1=1.,
c2=0.66,
)
self._agent_generator = sprite_generators.generate_sprites(
agent_factors, num_sprites=1)
# Predators
predator_factors = distribs.Product(
[
distribs.Continuous('x', 0., 1.),
distribs.Continuous('y', 0., 1.)
],
shape='circle',
scale=0.1,
c0=0.,
c1=1.,
c2=0.8,
)
self._predator_generator = sprite_generators.generate_sprites(
predator_factors, num_sprites=num_predators)
# Walls
self._walls = shapes.border_walls(visible_thickness=0.,
c0=0.,
c1=0.,
c2=0.5)
self._meta_state = None
def get_config(_):
"""Get environment config."""
############################################################################
# Sprite initialization
############################################################################
# Agent
agent_factors = distribs.Product(
[distribs.Continuous('x', 0.1, 0.9),
distribs.Continuous('y', 0.1, 0.9)],
shape='circle', scale=0.1, c0=0.33, c1=1., c2=0.66,
)
# Predators
shape_0 = 1.8 * np.array(
[[-0.3, -0.3], [0.1, -0.7], [0.4, 0.6], [-0.1, 0.25]])
shape_1 = 1.5 * np.array(
[[-0.5, -0.3], [-0.1, -0.7], [0.7, 0.1], [0., -0.1], [-0.3, 0.25]])
predator_factors = distribs.Product(
[distribs.Continuous('x', 0.2, 0.8),
distribs.Continuous('y', 0.2, 0.8),
distribs.Discrete(
'shape', [shape_0, shape_1, 'star_5', 'triangle', 'spoke_5']),
distribs.Continuous('angle', 0., 2 * np.pi),
distribs.Continuous('aspect_ratio', 0.75, 1.25),
distribs.Continuous('scale', 0.1, 0.15),
distribs.Continuous('x_vel', -0.03, 0.03),
distribs.Continuous('y_vel', -0.03, 0.03),
distribs.Continuous('angle_vel', -0.05, 0.05)],
c0=0., c1=1., c2=0.8,
)
# Walls
walls = shapes.border_walls(visible_thickness=0.05, c0=0., c1=0., c2=0.5)
# Create callable initializer returning entire state
agent_generator = sprite_generators.generate_sprites(
agent_factors, num_sprites=1)
predator_generator = sprite_generators.generate_sprites(
predator_factors, num_sprites=5)
def state_initializer():
predators = predator_generator(
disjoint=True, without_overlapping=walls)
agent = agent_generator(without_overlapping=walls + predators)
state = collections.OrderedDict([
('walls', walls),
('predators', predators),
('agent', agent),
])
return state
############################################################################
# Physics
############################################################################
agent_friction_force = physics_lib.Drag(coeff_friction=0.25)
asymmetric_collision = physics_lib.Collision(
elasticity=1., symmetric=False, update_angle_vel=True)
symmetric_collision = physics_lib.Collision(
elasticity=1., symmetric=True, update_angle_vel=True)
agent_wall_collision = physics_lib.Collision(
elasticity=0., symmetric=False, update_angle_vel=False)
forces = (
(agent_friction_force, 'agent'),
(symmetric_collision, 'predators', 'predators'),
(asymmetric_collision, 'predators', 'walls'),
(agent_wall_collision, 'agent', 'walls'),
)
physics = physics_lib.Physics(*forces, updates_per_env_step=10)
############################################################################
# Task
############################################################################
predator_task = tasks.ContactReward(
-5, layers_0='agent', layers_1='predators')
stay_alive_task = tasks.StayAlive(
reward_period=20,
reward_value=0.2,
)
task = tasks.CompositeTask(
predator_task, stay_alive_task, timeout_steps=200)
############################################################################
# Action space
############################################################################
action_space = action_spaces.Joystick(
scaling_factor=0.01, action_layers='agent')
############################################################################
# Observer
############################################################################
observer = observers.PILRenderer(
image_size=(64, 64), anti_aliasing=1, color_to_rgb='hsv_to_rgb')
############################################################################
# Final config
############################################################################
config = {
'state_initializer': state_initializer,
'physics': physics,
'task': task,
'action_space': action_space,
'observers': {'image': observer},
}
return config
def get_config(_):
"""Get environment config."""
############################################################################
# Sprite initialization
############################################################################
# Agent
agent_factors = distribs.Product(
[
distribs.Continuous('x', 0.1, 0.9),
distribs.Continuous('y', 0.1, 0.9)
],
shape='circle',
scale=0.1,
c0=0.33,
c1=1.,
c2=0.7,
)
# Predators
predator_factors = distribs.Product(
[
distribs.Continuous('x', 0.1, 0.9),
distribs.Continuous('y', 0.1, 0.9)
],
shape='circle',
scale=0.1,
c0=0.,
c1=1.,
c2=0.8,
)
# Prey
prey_factors = distribs.Product(
[
distribs.Continuous('x', 0.1, 0.9),
distribs.Continuous('y', 0.1, 0.9)
],
shape='circle',
scale=0.1,
c0=0.2,
c1=1.,
c2=1.,
)
# Boosters
booster_factors = distribs.Product(
[
distribs.Continuous('x', 0.1, 0.9),
distribs.Continuous('y', 0.1, 0.9)
],
shape='triangle',
scale=0.1,
c0=0.6,
c1=1.,
c2=1.,
)
# Portals
portal_factors = dict(shape='square', scale=0.1, c0=0., c1=0., c2=0.95)
portal_sprites = [
sprite.Sprite(x=0.125, y=0.125, **portal_factors),
sprite.Sprite(x=0.875, y=0.875, **portal_factors),
]
# Walls
wall_color = dict(c0=0., c1=0., c2=0.5)
island_wall_shape_0 = np.array([[0.2, 0.2], [0.4, 0.2], [0.4, 0.4],
[0.2, 0.4]])
island_wall_shapes = [
island_wall_shape_0,
island_wall_shape_0 + np.array([[0., 0.4]]),
island_wall_shape_0 + np.array([[0.4, 0.4]]),
island_wall_shape_0 + np.array([[0.4, 0.]]),
]
island_walls = [
sprite.Sprite(shape=shape, x=0., y=0., **wall_color)
for shape in island_wall_shapes
]
boundary_walls = shapes.border_walls(visible_thickness=0.05, **wall_color)
walls = boundary_walls + island_walls
# Callable sprite generators
agent_generator = sprite_generators.generate_sprites(agent_factors,
num_sprites=1)
predator_generator = sprite_generators.generate_sprites(predator_factors,
num_sprites=1)
prey_generator = sprite_generators.generate_sprites(
prey_factors, num_sprites=lambda: np.random.randint(2, 5))
booster_generator = sprite_generators.generate_sprites(booster_factors,
num_sprites=2)
# Create callable initializer returning entire state
def state_initializer():
portals = portal_sprites
agent = agent_generator(without_overlapping=walls)
predators = predator_generator(without_overlapping=walls + agent)
boosters = booster_generator(without_overlapping=walls + agent)
prey = prey_generator(without_overlapping=walls)
state = collections.OrderedDict([
('walls', walls),
('portals', portals),
('boosters', boosters),
('prey', prey),
('predators', predators),
('agent', agent),
])
return state
############################################################################
# Physics
############################################################################
agent_friction_force = physics_lib.Drag(coeff_friction=0.25)
predator_friction_force = physics_lib.Drag(coeff_friction=0.05)
predator_random_force = physics_lib.RandomForce(max_force_magnitude=0.02)
prey_friction_force = physics_lib.Drag(coeff_friction=0.02)
prey_random_force = physics_lib.RandomForce(max_force_magnitude=0.02)
predator_attraction = physics_lib.DistanceForce(
force_fn=physics_lib.linear_force_fn(zero_intercept=-0.002,
slope=0.001))
asymmetric_collision = physics_lib.Collision(elasticity=0.25,
symmetric=False,
update_angle_vel=False)
forces = (
(agent_friction_force, 'agent'),
(predator_friction_force, 'predators'),
(predator_random_force, 'predators'),
(prey_friction_force, 'prey'),
(prey_random_force, 'prey'),
(predator_attraction, 'agent', 'predators'),
(asymmetric_collision, ['agent', 'predators', 'prey'], 'walls'),
)
physics = physics_lib.Physics(*forces, updates_per_env_step=5)
############################################################################
# Task
############################################################################
predator_task = tasks.ContactReward(-5,
layers_0='agent',
layers_1='predators',
reset_steps_after_contact=0)
prey_task = tasks.ContactReward(1, layers_0='agent', layers_1='prey')
reset_task = tasks.Reset(
condition=lambda state: len(state['prey']) == 0,
steps_after_condition=5,
)
task = tasks.CompositeTask(predator_task,
prey_task,
reset_task,
timeout_steps=400)
############################################################################
# Action space
############################################################################
action_space = action_spaces.Joystick(
scaling_factor=0.01,
action_layers='agent',
)
############################################################################
# Observer
############################################################################
observer = observers.PILRenderer(
image_size=(64, 64),
anti_aliasing=1,
color_to_rgb='hsv_to_rgb',
)
############################################################################
# Game rules
############################################################################
disappear_rule = game_rules.VanishOnContact(vanishing_layer='prey',
contacting_layer='agent')
portal_rule = game_rules.Portal(teleporting_layer='agent',
portal_layer='portals')
rules = (disappear_rule, portal_rule, Booster())
############################################################################
# Final config
############################################################################
config = {
'state_initializer': state_initializer,
'physics': physics,
'task': task,
'action_space': action_space,
'observers': {
'image': observer
},
'game_rules': rules,
}
return config
def get_config(_):
"""Get environment config."""
############################################################################
# Sprite initialization
############################################################################
# Agents
agent_factors = distribs.Product(
[distribs.Continuous('x', 0., 1.),
distribs.Continuous('y', 0.35, 0.65)],
shape='circle', scale=0.1, c1=1., c2=0.7,
)
agent_0_factors = distribs.Product([agent_factors], c0=0.2)
agent_1_factors = distribs.Product([agent_factors], c0=0.1)
agent_2_factors = distribs.Product([agent_factors], c0=0.)
# Walls
walls = shapes.border_walls(visible_thickness=0.05, c0=0., c1=0., c2=0.5)
# Fountains
fountain_factors = {
'shape': 'circle', 'scale': 0.05, 'c0': 0.6, 'c1': 1., 'c2': _BAD_VALUE}
fountains_across = np.linspace(0.1, 0.9, 6)
fountains_up = np.linspace(0.75, 0.9, 2)
fountains_grid_x, fountains_grid_y = np.meshgrid(fountains_across,
fountains_up)
fountains_positions = zip(np.ravel(fountains_grid_x),
np.ravel(fountains_grid_y))
fountain_sprites = [
sprite.Sprite(x=x, y=y, **fountain_factors)
for (x, y) in fountains_positions
]
# Fruits
fruit_factors = {
'shape': 'circle', 'scale': 0.05, 'c0': 0.3, 'c1': 1., 'c2': _BAD_VALUE}
fruits_across = np.linspace(0.1, 0.9, 6)
fruits_up = np.linspace(0.1, 0.25, 2)
fruits_grid_x, fruits_grid_y = np.meshgrid(fruits_across, fruits_up)
fruits_positions = zip(np.ravel(fruits_grid_x), np.ravel(fruits_grid_y))
fruit_sprites = [
sprite.Sprite(x=x, y=y, **fruit_factors)
for (x, y) in fruits_positions
]
# Create callable initializer returning entire state
agent_0_generator = sprite_generators.generate_sprites(
agent_0_factors, num_sprites=1)
agent_1_generator = sprite_generators.generate_sprites(
agent_1_factors, num_sprites=1)
agent_2_generator = sprite_generators.generate_sprites(
agent_2_factors, num_sprites=1)
def state_initializer():
agent_0 = agent_0_generator(without_overlapping=walls)
agent_1 = agent_1_generator(without_overlapping=walls)
agent_2 = agent_2_generator(without_overlapping=walls)
state = collections.OrderedDict([
('walls', walls),
('fountains', fountain_sprites),
('fruits', fruit_sprites),
('agent_2', agent_2),
('agent_1', agent_1),
('agent_0', agent_0),
])
return state
############################################################################
# Physics
############################################################################
agent_friction_force = physics_lib.Drag(coeff_friction=0.25)
asymmetric_collision = physics_lib.Collision(
elasticity=0.25, symmetric=False)
forces = (
(agent_friction_force, ['agent_0', 'agent_1', 'agent_2']),
(asymmetric_collision, ['agent_0', 'agent_1', 'agent_2'], 'walls'),
)
physics = physics_lib.Physics(*forces, updates_per_env_step=5)
############################################################################
# Task
############################################################################
task = tasks.ContactReward(
1, layers_0='agent_0', layers_1='fruits',
condition=lambda s_0, s_1: s_1.c2 > _VALUE_THRESHOLD,
)
############################################################################
# Action space
############################################################################
action_space = action_spaces.Composite(
agent_0=action_spaces.Joystick(
scaling_factor=0.005, action_layers='agent_0'),
agent_1=action_spaces.Joystick(
scaling_factor=0.005, action_layers='agent_1'),
agent_2=action_spaces.Joystick(
scaling_factor=0.005, action_layers='agent_2'),
)
############################################################################
# Observer
############################################################################
image_observer = observers.PILRenderer(
image_size=(64, 64),
anti_aliasing=1,
color_to_rgb='hsv_to_rgb',
)
raw_state_observer = observers.RawState() # needed by hand-crafted agents
############################################################################
# Game rules
############################################################################
def _spoil_fruit(sprite):
sprite.c2 = _BAD_VALUE
def _ripen_fruit(sprite):
sprite.c2 = _GOOD_VALUE
def _poison_fountain(sprite):
sprite.c2 = _BAD_VALUE
def _clean_fountain(sprite):
sprite.c2 = _GOOD_VALUE
def agents_contacting_layer(state, layer, value):
n_contact = 0
for s in state[layer]:
if s.c2 != value:
continue
n_contact += (
s.overlaps_sprite(state['agent_0'][0]) or
s.overlaps_sprite(state['agent_1'][0]) or
s.overlaps_sprite(state['agent_2'][0])
)
return n_contact
poison_fountains = game_rules.ModifySprites(
layers='fountains', modifier=_poison_fountain, sample_one=True,
filter_fn=lambda s: s.c2 > _VALUE_THRESHOLD)
poison_fountains = game_rules.ConditionalRule(
condition=lambda s: agents_contacting_layer(s, 'fruits', _GOOD_VALUE),
rules=poison_fountains,
)
ripen_fruits = game_rules.ModifySprites(
layers='fruits', modifier=_ripen_fruit, sample_one=True,
filter_fn=lambda s: s.c2 < _VALUE_THRESHOLD)
ripen_fruits = game_rules.ConditionalRule(
condition=lambda s: agents_contacting_layer(s, 'fountains', _BAD_VALUE),
rules=ripen_fruits,
)
spoil_fruits = game_rules.ModifyOnContact(
layers_0='fruits',
layers_1=('agent_0', 'agent_1', 'agent_2'),
modifier_0=_spoil_fruit,
filter_0=lambda s: s.c2 > _VALUE_THRESHOLD)
clean_fountains = game_rules.ModifyOnContact(
layers_0='fountains',
layers_1=('agent_0', 'agent_1', 'agent_2'),
modifier_0=_clean_fountain,
filter_0=lambda s: s.c2 < _VALUE_THRESHOLD)
rules = (poison_fountains, spoil_fruits, ripen_fruits, clean_fountains)
############################################################################
# Final config
############################################################################
config = {
'state_initializer': state_initializer,
'physics': physics,
'task': task,
'action_space': action_space,
'observers': {'image': image_observer, 'state': raw_state_observer},
'game_rules': rules,
}
return config