def __init__(self,
*,
num_agents,
num_hazards,
identical_rewards,
observation_type,
hazard_risk=0.5,
collision_risk=0.0):
'''
Args:
- num_agents [int] number of agents in scenario
- num_hazards [int] number of hazards landmarks in the scenario
- identical_rewards [bool] true if all agents receieve exact same reward, false if rewards are "local" to agents
- observation_type [str] "direct" if observation directly of each entity, "histogram" if bin entities in spacial grid
- hazard_risk [float] max probability of failure caused by hazard landmark
- collision_risk [float] probability of failure caused by collision
'''
# check inputs
assert isinstance(num_agents, int)
assert isinstance(num_hazards, int)
assert (num_hazards == 0 or num_hazards == 1)
assert isinstance(identical_rewards, bool)
assert (observation_type == "direct"
or observation_type == "histogram")
assert (hazard_risk >= 0.0 and hazard_risk <= 1.0)
assert (collision_risk >= 0.0 and collision_risk <= 1.0)
# set member vars
self.num_agents = num_agents
self.num_hazards = num_hazards
self.identical_rewards = identical_rewards
self.observation_type = observation_type
self.hazard_risk = hazard_risk
self.collision_risk = collision_risk
# create list of landmarks
# Note: RadialReward function is not directly used for calculating reward in this scenario, thus peak value of 0.0.
# non-zero radius used for numerical reasons
landmarks = []
for i in range(self.num_agents):
landmarks.append(
RiskRewardLandmark(risk_fn=None,
reward_fn=RadialReward(1.0, 0.0)))
for i in range(self.num_hazards):
landmarks.append(
RiskRewardLandmark(risk_fn=RadialRisk(_LANDMARK_SIZE, 0.5),
reward_fn=RadialReward(1.0, 0.0)))
self.scenario_landmarks = landmarks
self.n_landmarks = len(self.scenario_landmarks)
def make_world(self):
world = HazardousWorld()
# set scenario-specific world parameters
world.collaborative = True
world.systemic_rewards = True
world.identical_rewards = False
world.dim_c = 0 # observation-based communication
world.max_communication_distance = _MAX_COMMUNICATION_DISTANCE
world.distance_resistance_gain = _DISTANCE_RESISTANCE_GAIN
world.distance_resistance_exponent = _DISTANCE_RESISTANCE_EXPONENT
# add landmarks
world.origin_terminal_landmark = RiskRewardLandmark( risk_fn=None, reward_fn=RadialReward(1.0, 10.0))
world.destination_terminal_landmark = RiskRewardLandmark( risk_fn=None, reward_fn=RadialReward(1.0, 10.0))
world.landmarks = [world.origin_terminal_landmark, world.destination_terminal_landmark]
for lm in _NON_TERMINAL_LANDMARKS:
world.landmarks.append(lm)
for i, landmark in enumerate(world.landmarks):
landmark.name = 'landmark_%d' % i
landmark.collide = False
landmark.movable = False
landmark.size = _LANDMARK_SIZE
# properties for landmarks
if isinstance(landmark, RiskRewardLandmark) and landmark.is_hazard:
#TODO: make colors heatmap of risk probability over all bounds
landmark.color = np.array([landmark.risk_fn.get_failure_probability(0,0) + .1, 0, 0])
else:
landmark.color = np.array([0.25, 0.25, 0.25])
# make initial conditions
self.reset_world(world)
return world
def make_world(self):
world = HazardousWorld()
# set scenario-specific world parameters
world.collaborative = True
world.systemic_rewards = True
world.identical_rewards = False
world.dim_c = 0 # observation-based communication
world.connection_reward = _CONNECTION_REWARD
world.termination_reward = _TERMINATION_REWARD
world.render_connections = True
# add landmarks
world.origin_terminal_landmark = RiskRewardLandmark(
risk_fn=None, reward_fn=RadialReward(1.0, 10.0))
world.destination_terminal_landmark = RiskRewardLandmark(
risk_fn=None, reward_fn=RadialReward(1.0, 10.0))
world.hazard_landmark = _NON_TERMINAL_LANDMARKS[0]
world.hazard_landmark.ignore_connection_rendering = True
world.landmarks = [
world.origin_terminal_landmark,
world.destination_terminal_landmark, world.hazard_landmark
]
for i, landmark in enumerate(world.landmarks):
landmark.name = 'landmark_%d' % i
landmark.collide = False
landmark.movable = False
landmark.size = _LANDMARK_SIZE
# properties for landmarks
if isinstance(landmark, RiskRewardLandmark) and landmark.is_hazard:
#TODO: make colors heatmap of risk probability over all bounds
landmark.color = np.array([
landmark.risk_fn.get_failure_probability(0, 0) + .1, 0, 0
])
else:
landmark.color = np.array([0.25, 0.25, 0.25])
# make initial conditions
self.reset_world(world)
return world
def make_world(self):
world = HazardousWorld()
# set scenario-specific world parameters
world.collaborative = True
world.systemic_rewards = True
world.identical_rewards = False
world.dim_c = 0 # observation-based communication
world.max_communication_distance = _MAX_COMMUNICATION_DISTANCE
world.distance_resistance_gain = _DISTANCE_RESISTANCE_GAIN
# create and add terminal landmarks
# no intermediate landmarks in this scenario
world.origin_terminal_landmark = RiskRewardLandmark(
risk_fn=None, reward_fn=RadialReward(1.0, 10.0))
world.origin_terminal_landmark.name = 'origin'
world.origin_terminal_landmark.state.p_pos = np.array([-0.75, -0.75])
world.destination_terminal_landmark = RiskRewardLandmark(
risk_fn=None, reward_fn=RadialReward(1.0, 10.0))
world.destination_terminal_landmark.name = 'destination'
world.destination_terminal_landmark.state.p_pos = np.array(
[0.75, 0.75])
# create landmark list and set properties
world.landmarks = [
world.origin_terminal_landmark, world.destination_terminal_landmark
]
for i, landmark in enumerate(world.landmarks):
landmark.p_vel = np.zeros(world.dim_p)
landmark.collide = False
landmark.movable = False
landmark.size = _LANDMARK_SIZE
landmark.color = np.array(
[landmark.risk_fn.get_failure_probability(0, 0) + .1, 0, 0])
# make initial conditions
self.reset_world(world)
return world
from particle_environments.common import RadialPolynomialRewardFunction2D as RadialReward
from particle_environments.common import RadialBernoulliRiskFunction2D as RadialRisk
from rl_algorithms.scenariolearning import ScenarioHeuristicAgentTrainer
# Scenario Parameters
_MAX_COMMUNICATION_DISTANCE = np.inf
_AGENT_SIZE = 0.15
_LANDMARK_SIZE = 0.05
_AGENT_OBSERVATION_LEN = 5
_LANDMARK_OBSERVATION_LEN = 3
_NUM_AGENTS = 3
_LANDMARKS = []
_LANDMARKS.append(
RiskRewardLandmark(risk_fn=RadialRisk(_LANDMARK_SIZE, 0.5),
reward_fn=RadialReward(1.0, 0.0)))
_LANDMARKS.append(
RiskRewardLandmark(risk_fn=None, reward_fn=RadialReward(1.0, 0.0)))
_LANDMARKS.append(
RiskRewardLandmark(risk_fn=None, reward_fn=RadialReward(1.0, 0.0)))
_LANDMARKS.append(
RiskRewardLandmark(risk_fn=None, reward_fn=RadialReward(1.0, 0.0)))
_N_LANDMARKS = len(_LANDMARKS)
class Scenario(BaseScenario):
# static class
num_agents = _NUM_AGENTS
def make_world(self):
world = HazardousWorld(collision_termination_probability=0.0)
_MAX_CONNECTION_DISTANCE = 0.35
_MAX_OBSERVATION_DISTANCE = 1.0
_CONNECTION_REWARD = 1.0
_TERMINATION_REWARD = -0.0
_AGENT_SIZE = 0.01
_LANDMARK_SIZE = 0.025
_N_RADIAL_BINS = 4
_N_ANGULAR_BINS = 8
_N_OBSERVED_TERMINATIONS = 5
_N_TERMINALS = 2
_ZERO_THRESHOLD = 1e-6
_NON_TERMINAL_LANDMARKS = []
_NON_TERMINAL_LANDMARKS.append(
RiskRewardLandmark(risk_fn=RadialRisk(0.1),
reward_fn=RadialReward(0.1, 10.0)))
class Scenario(BaseScenario):
# static class
num_agents = 20
def make_world(self):
world = HazardousWorld()
# set scenario-specific world parameters
world.collaborative = True
world.systemic_rewards = True
world.identical_rewards = False
world.dim_c = 0 # observation-based communication
world.connection_reward = _CONNECTION_REWARD
from particle_environments.mager.world import TemporarilyObservableRiskRewardLandmark as TORRLandmark
from particle_environments.mager.observation import format_observation
from particle_environments.common import is_collision, distance, delta_pos
from particle_environments.common import RadialPolynomialRewardFunction2D as RadialReward
from particle_environments.common import RadialBernoulliRiskFunction2D as RadialRisk
from particle_environments.common import DefaultParameters as DP
# Scenario Parameters
_MAX_COMMUNICATION_DISTANCE = 0.5
_AGENT_SIZE = 0.01
_LANDMARK_SIZE = 0.025
_LANDMARKS = []
_LANDMARKS.append(
TORRLandmark(risk_fn=RadialRisk(0.1),
reward_fn=RadialReward(0.15, 10.0),
observe_duration=1.0))
_LANDMARKS.append(
TORRLandmark(risk_fn=RadialRisk(0.1),
reward_fn=RadialReward(0.15, 10.0),
observe_duration=1.0))
_LANDMARKS.append(
TORRLandmark(risk_fn=RadialRisk(0.1),
reward_fn=RadialReward(0.15, 10.0),
observe_duration=1.0))
class Scenario(BaseScenario):
# static class
num_agents = 10