Note
Click here to download the full example code or to run this example in your browser via Binder
Reinforcement Learning Sensor Manager
This example looks at how to interface a reinforcement learning framework with a Stone Soup sensor manager.
Making a Reinforcement Learning Sensor Manager
This example introduces using a Deep Q Network (DQN) reinforcement learning (RL) sensor management algorithm
in Stone Soup. This is compared to the performance of a brute force algorithm using the same metrics shown in the
sensor management tutorials. This example is similar to the sensor management tutorials, simulating 3 targets and a
RadarRotatingBearingRange sensor which can be actioned to point in different directions.
Tensorflow-agents is used as the reinforcement learning framework. This is a separate python package that can be found at https://github.com/tensorflow/agents.
Warning
This currently only works on Linux based OSes, or via Windows Subsystem for Linux (WSL). MacOS users may be able to make use of a Linux VM to run this example. See Tensorflow instructions for creating Python virtual environments (with GPU support if applicable) [1].
To run this example, in a clean environment, do pip install stonesoup, followed by pip install
tf-agents[reverb].
# Some general imports and set up
import numpy as np
import random
from datetime import datetime, timedelta
try:
import reverb
except ImportError:
raise ImportError('To run this example, reverb must be installed. Please read the warning'
'and instructions at the top of this notebook.')
start_time = datetime.now().replace(microsecond=0)
from stonesoup.models.transition.linear import CombinedLinearGaussianTransitionModel, ConstantVelocity
from stonesoup.types.groundtruth import GroundTruthPath, GroundTruthState
Generate ground truths
Following the methods from previous Stone Soup sensor management tutorials, generate a series of combined linear Gaussian transition models and generate ground truths. Each ground truth is offset in the y-direction by 10.
The number of targets in this simulation is defined by ntruths - here there are 3 targets travelling in different directions. The time the simulation is observed for is defined by time_max.
We can fix our random number generator in order to probe a particular example repeatedly. To produce random examples, comment out the next two lines.
np.random.seed(1990)
random.seed(1990)
# Generate transition model
# i.e. fk(xk|xk-1)
transition_model = CombinedLinearGaussianTransitionModel([ConstantVelocity(0.005),
ConstantVelocity(0.005)])
yps = range(0, 100, 10) # y value for prior state
truths = []
ntruths = 3 # number of ground truths in simulation
time_max = 50 # timestamps the simulation is observed over
timesteps = [start_time + timedelta(seconds=k) for k in range(time_max)]
xdirection = 1
ydirection = 1
# Generate ground truths
for j in range(0, ntruths):
truth = GroundTruthPath([GroundTruthState([0, xdirection, yps[j], ydirection],
timestamp=start_time)],
id=f"id{j}")
for k in range(1, time_max):
truth.append(
GroundTruthState(transition_model.function(truth[k - 1], noise=True, time_interval=timedelta(seconds=1)),
timestamp=start_time + timedelta(seconds=k)))
truths.append(truth)
# alternate directions when initiating tracks
xdirection *= -1
if j % 2 == 0:
ydirection *= -1
Plot the ground truths. This is done using the Plotterly class from Stone Soup.
from stonesoup.plotter import AnimatedPlotterly
plotter = AnimatedPlotterly(timesteps, tail_length=1)
plotter.plot_ground_truths(truths, [0, 2])
plotter.fig
Create sensors
Create a sensor for each sensor management algorithm. This tutorial uses the
RadarRotatingBearingRange sensor. This sensor is an Actionable so
is capable of returning the actions it can take at a given time step and can also be given an action to take before
measuring.
See the Creating an Actionable Sensor Example for a more
detailed explanation of actionable sensors.
The RadarRotatingBearingRange has a dwell centre which is an ActionableProperty
so in this case the action is changing the dwell centre to point in a specific direction.
from stonesoup.types.state import StateVector
from stonesoup.sensor.radar.radar import RadarRotatingBearingRange
sensorA = RadarRotatingBearingRange(
position_mapping=(0, 2),
noise_covar=np.array([[np.radians(0.5) ** 2, 0],
[0, 1 ** 2]]),
ndim_state=4,
position=np.array([[10], [0]]),
rpm=60,
fov_angle=np.radians(45),
dwell_centre=StateVector([0.0]),
max_range=np.inf
)
sensorA.timestamp = start_time
sensorB = RadarRotatingBearingRange(
position_mapping=(0, 2),
noise_covar=np.array([[np.radians(0.5) ** 2, 0],
[0, 1 ** 2]]),
ndim_state=4,
position=np.array([[10], [0]]),
rpm=60,
fov_angle=np.radians(45),
dwell_centre=StateVector([0.0]),
max_range=np.inf
)
sensorB.timestamp = start_time
Create the Kalman predictor and updater
Construct a predictor and updater using the KalmanPredictor and ExtendedKalmanUpdater
components from Stone Soup. The ExtendedKalmanUpdater is used because it can be used for both linear
and nonlinear measurement models. A hypothesiser and data associator are required for use in both trackers.
from stonesoup.predictor.kalman import KalmanPredictor
predictor = KalmanPredictor(transition_model)
from stonesoup.updater.kalman import ExtendedKalmanUpdater
updater = ExtendedKalmanUpdater(measurement_model=None)
# measurement model is added to detections by the sensor
from stonesoup.hypothesiser.distance import DistanceHypothesiser
from stonesoup.measures import Mahalanobis
hypothesiser = DistanceHypothesiser(predictor, updater, measure=Mahalanobis(), missed_distance=5)
from stonesoup.dataassociator.neighbour import GNNWith2DAssignment
data_associator = GNNWith2DAssignment(hypothesiser)
Generate Priors
First create ntruths priors which estimate the targets’ initial states, one for each target. In this example each prior is offset by 0.1 in the y direction meaning the position of the track is initially not very accurate. The velocity is also systematically offset by +0.2 in both the x and y directions.
from stonesoup.types.state import GaussianState
priors = []
xdirection = 1.2
ydirection = 1.2
for j in range(0, ntruths):
priors.append(GaussianState([[0], [xdirection], [yps[j] + 0.1], [ydirection]],
np.diag([0.5, 0.5, 0.5, 0.5] + np.random.normal(0, 5e-4, 4)),
timestamp=start_time))
xdirection *= -1
if j % 2 == 0:
ydirection *= -1
Initialise the tracks by creating an empty list and appending the priors generated. This needs to be done separately for both sensor manager methods as they will generate different sets of tracks.
from stonesoup.types.track import Track
# Initialise tracks from the RandomSensorManager
tracksA = []
for j, prior in enumerate(priors):
tracksA.append(Track([prior]))
tracksB = []
for j, prior in enumerate(priors):
tracksB.append(Track([prior]))
Reward function
A reward function is used to quantify the benefit of sensors taking a particular action or set of actions. This can be crafted specifically for an example in order to achieve a particular objective. The function used in this example is quite generic but could be substituted for any callable function which returns a numeric value that the sensor manager can maximise.
The UncertaintyRewardFunction calculates the uncertainty reduction by computing the difference between the
covariance matrix norms of the prediction, and the posterior assuming a predicted measurement corresponding to that
prediction.
from stonesoup.sensormanager.reward import UncertaintyRewardFunction
reward_function = UncertaintyRewardFunction(predictor=predictor, updater=updater)
Reinforcement Learning
Reinforcement learning involves intelligent agents making decisions to maximise a cumulative reward. The agent must train in an environment in order to create a policy, which later determines the actions it will take. During training, the agent makes decisions and receives rewards, which it uses to optimise the policy.
Illustration of an RL algorithm taking actions during training. The state and reward it receives are used to determine the best actions.
Once training has completed, the policy can be exploited to gain rewards.
Design Environment
An environment is needed for the RL agent to learn in. There are resources online for how to design these [2].
In this example, the action space is equal to the number of targets in the simulation, so at each time step, the
sensor can select one target to look at. For the environment, we make a copy of the sensor that we will pass to the
sensor manager later on. This is so the agent can train in the environment without altering the sensor itself.
The UncertaintyRewardFunction is used to calculate the reward obtained for each step in the environment.
The trace of the covariances for each object is used as the observation for the agent to learn from - it should learn
to select targets with a larger covariance (higher uncertainty).
from abc import ABC
import numpy as np
import copy
from ordered_set import OrderedSet
from stonesoup.sensor.action.dwell_action import DwellActionsGenerator
from stonesoup.functions import mod_bearing
from tf_agents.environments import py_environment
from tf_agents.specs import array_spec
from tf_agents.trajectories import time_step as ts
from tf_agents.environments import utils
class StoneSoupEnv(py_environment.PyEnvironment, ABC):
"""Example reinforcement learning environment. Environments must contain __init__, _reset,
_step, and generate_action methods
"""
def __init__(self):
super().__init__()
# Action size is number of targets
self._action_spec = array_spec.BoundedArraySpec(
shape=(), dtype=np.int32, minimum=0, maximum=ntruths - 1, name='action')
# Observation size is also number of targets
self.obs_size = ntruths
self._observation_spec = array_spec.BoundedArraySpec(
shape=(self.obs_size,), dtype=np.float32, name='observation')
self._episode_ended = False
self.max_episode_length = time_max
self.current_step = 0
self.start_time = start_time
# Use deepcopy to prevent the original sensor/tracks being changed each time an episode is run
self.sensor = copy.deepcopy(sensorA)
self.sensor.timestamp = start_time
self.tracks = copy.deepcopy(tracksA)
def action_spec(self):
"""Return action_spec."""
return self._action_spec
def observation_spec(self):
"""Return observation_spec."""
return self._observation_spec
def _reset(self):
"""Restarts the environment from the first step, resets the initial state
and observation values, and returns an initial observation
"""
self._episode_ended = False
self.current_step = 0
self.sensor = copy.deepcopy(sensorA)
self.sensor.timestamp = start_time
self.tracks = copy.deepcopy(tracksA)
return ts.restart(np.zeros((self.obs_size,), dtype=np.float32))
def _step(self, action):
"""Apply action and take one step through environment, and return new time_step.
"""
reward = 0
if self._episode_ended:
# The last action ended the episode. Ignore the current action and start
# a new episode.
return self.reset()
uncertainty = []
for i, target in enumerate(self.tracks):
# Calculate the bearing of the chosen target from the sensor
if i == action:
x_target = target.state.state_vector[0] - self.sensor.position[0]
y_target = target.state.state_vector[2] - self.sensor.position[1]
bearing_target = mod_bearing(np.arctan2(y_target, x_target))
uncertainty.append(np.trace(target.covar))
current_timestep = self.start_time + timedelta(seconds=self.current_step)
next_timestep = self.start_time + timedelta(seconds=self.current_step + 1)
# Create action generator which contains possible actions
action_generator = DwellActionsGenerator(self.sensor,
attribute='dwell_centre',
start_time=current_timestep,
end_time=next_timestep)
# Action the environment's sensor to point towards the chosen target
current_action = [action_generator.action_from_value(bearing_target)]
config = ({self.sensor: current_action})
reward += reward_function(config, self.tracks, next_timestep)
self.sensor.add_actions(current_action)
self.sensor.act(next_timestep)
# Calculate a measurement from the sensor
measurement = set()
measurement |= self.sensor.measure(OrderedSet(truth[current_timestep] for truth in truths), noise=True)
hypotheses = data_associator.associate(self.tracks,
measurement,
current_timestep)
for track in self.tracks:
hypothesis = hypotheses[track]
if hypothesis.measurement:
post = updater.update(hypothesis)
track.append(post)
else: # When data associator says no detections are good enough, we'll keep the prediction
track.append(hypothesis.prediction)
# Set the observation as the prior uncertainty of each target
observation = np.array(uncertainty, dtype=np.float32)
self.current_step += 1
if self.current_step >= self.max_episode_length - 1:
self._episode_ended = True
return ts.termination(observation, reward)
else:
return ts.transition(observation, reward=reward, discount=1.0)
@staticmethod
def generate_action(action, tracks, sensor):
"""This method is used to convert a tf-agents action into a Stone Soup action"""
for i, target in enumerate(tracks):
if i == action:
x_target = target.state.state_vector[0] - sensor.position[0]
y_target = target.state.state_vector[2] - sensor.position[1]
action_bearing = mod_bearing(np.arctan2(y_target, x_target))
action_generators = DwellActionsGenerator(sensor,
attribute='dwell_centre',
start_time=sensor.timestamp,
end_time=sensor.timestamp + timedelta(seconds=1))
current_action = [action_generators.action_from_value(action_bearing)]
return current_action
# Validate the environment to ensure that the environment returns the expected specs
train_env = StoneSoupEnv()
utils.validate_py_environment(train_env, episodes=5)
RL Sensor Manager
To be able to use the RL environment we have designed, we need to make a ReinforcementSensorManager class, which
inherits from SensorManager.
We introduce some additional methods that are used by tensorflow-agents: compute_avg_return(),
dense_layer(), and train().
compute_avg_return() is used to find the average reward by using a given policy. This is used to evaluate the
training.
dense_layer() is used when generating the Q-Network, a neural network model that learns to predict Q-Values.
train() is used to generate the policy by running a large number of episodes through the Q-Network to work out
which actions are best. An episode in RL refers to a single run or instance of the learning process, where the agent
interacts with the environment.
We also need to re-define the choose_actions() method from SensorManager to be able to interface
Stone Soup actions with tensorflow-agent actions.
from stonesoup.sensormanager.base import SensorManager
from stonesoup.base import Property
from tf_agents.environments import tf_py_environment
class ReinforcementSensorManager(SensorManager):
"""A sensor manager that employs reinforcement learning algorithms from tensorflow-agents.
The sensor manager trains on an environment to find an optimal policy, which is then exploited
to choose actions.
"""
env: py_environment.PyEnvironment = Property(doc="The environment which the agent learns the policy with.")
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.tf_env = tf_py_environment.TFPyEnvironment(self.env)
self.test_env = tf_py_environment.TFPyEnvironment(self.env)
self.agent = None
@staticmethod
def compute_avg_return(environment, policy, num_episodes=10):
"""Used to calculate the average reward over a set of episodes.
Parameters
----------
environment:
tf-agents environment for evaluating policy on
policy:
tf-agents policy for choosing actions in environment
num_episodes: int
Number of episodes to sample over
Returns
-------
: int
average reward calculated over num_episodes
"""
time_step = None
episode_return = None
total_return = 0.0
for _ in range(num_episodes):
time_step = environment.reset()
episode_return = 0.0
while not time_step.is_last():
action_step = policy.action(time_step)
time_step = environment.step(action_step.action)
episode_return += time_step.reward
total_return += episode_return
avg_return = total_return / num_episodes
return avg_return.numpy()[0]
@staticmethod
def dense_layer(num_units):
"""Method for generating fully connected layers for use in the neural network.
Parameters
----------
num_units: int
Number of nodes in dense layer
Returns
-------
: tensorflow dense layer
"""
# Define a helper function to create Dense layers configured with the right
# activation and kernel initializer.
return tf.keras.layers.Dense(
num_units,
activation=tf.keras.activations.relu,
kernel_initializer=tf.keras.initializers.VarianceScaling(
scale=2.0, mode='fan_in', distribution='truncated_normal'))
def train(self, hyper_parameters):
"""Trains a DQN agent on the specified environment to learn a policy that is later
used to select actions.
Parameters
----------
hyper_parameters: dict
Dictionary containing hyperparameters used in training. See tutorial for
necessary hyperparameters.
"""
if self.env is not None:
self.env.reset()
train_py_env = self.env
eval_py_env = self.env
self.train_env = tf_py_environment.TFPyEnvironment(train_py_env)
self.eval_env = tf_py_environment.TFPyEnvironment(eval_py_env)
fc_layer_params = hyper_parameters['fc_layer_params']
action_tensor_spec = tensor_spec.from_spec(self.env.action_spec())
num_actions = action_tensor_spec.maximum - action_tensor_spec.minimum + 1
# QNetwork consists of a sequence of Dense layers followed by a dense layer
# with `num_actions` units to generate one q_value per available action as
# its output.
dense_layers = [self.dense_layer(num_units) for num_units in fc_layer_params]
q_values_layer = tf.keras.layers.Dense(
num_actions,
activation=None,
kernel_initializer=tf.keras.initializers.RandomUniform(
minval=-0.03, maxval=0.03),
bias_initializer=tf.keras.initializers.Constant(-0.2))
q_net = sequential.Sequential(dense_layers + [q_values_layer])
optimizer = tf.keras.optimizers.Adam(hyper_parameters['learning_rate'])
train_step_counter = tf.Variable(0)
self.agent = dqn_agent.DdqnAgent(
self.train_env.time_step_spec(),
self.train_env.action_spec(),
q_network=q_net,
optimizer=optimizer,
td_errors_loss_fn=common.element_wise_squared_loss,
train_step_counter=train_step_counter)
self.agent.initialize()
random_policy = random_tf_policy.RandomTFPolicy(self.train_env.time_step_spec(),
self.train_env.action_spec())
# See also the metrics module for standard implementations of different metrics.
# https://github.com/tensorflow/agents/tree/master/tf_agents/metrics
self.compute_avg_return(self.eval_env, random_policy,
hyper_parameters['num_eval_episodes'])
table_name = 'uniform_table'
replay_buffer_signature = tensor_spec.from_spec(
self.agent.collect_data_spec)
replay_buffer_signature = tensor_spec.add_outer_dim(
replay_buffer_signature)
table = reverb.Table(
table_name,
max_size=hyper_parameters['replay_buffer_max_length'],
sampler=reverb.selectors.Uniform(),
remover=reverb.selectors.Fifo(),
rate_limiter=reverb.rate_limiters.MinSize(1),
signature=replay_buffer_signature)
reverb_server = reverb.Server([table])
replay_buffer = reverb_replay_buffer.ReverbReplayBuffer(
self.agent.collect_data_spec,
table_name=table_name,
sequence_length=2,
local_server=reverb_server)
rb_observer = reverb_utils.ReverbAddTrajectoryObserver(
replay_buffer.py_client,
table_name,
sequence_length=2)
py_driver.PyDriver(
self.env,
py_tf_eager_policy.PyTFEagerPolicy(
random_policy, use_tf_function=True),
[rb_observer],
max_steps=hyper_parameters['initial_collect_steps']).run(train_py_env.reset())
# Dataset generates trajectories with shape [Bx2x...]
dataset = replay_buffer.as_dataset(
num_parallel_calls=3,
sample_batch_size=hyper_parameters['batch_size'],
num_steps=2).prefetch(3)
iterator = iter(dataset)
# (Optional) Optimize by wrapping some code in a graph using TF function.
self.agent.train = common.function(self.agent.train)
# Reset the train step.
self.agent.train_step_counter.assign(0)
# Evaluate the agent's policy once before training.
avg_return = self.compute_avg_return(self.eval_env, self.agent.policy,
hyper_parameters['num_eval_episodes'])
returns = [avg_return]
# Reset the environment.
time_step = train_py_env.reset()
# Create a driver to collect experience.
collect_driver = py_driver.PyDriver(
self.env,
py_tf_eager_policy.PyTFEagerPolicy(
self.agent.collect_policy, use_tf_function=True),
[rb_observer],
max_steps=hyper_parameters['collect_steps_per_iteration'])
for _ in range(hyper_parameters['num_iterations']):
# Collect a few steps and save to the replay buffer.
time_step, _ = collect_driver.run(time_step)
# Sample a batch of data from the buffer and update the agent's network.
experience, unused_info = next(iterator)
train_loss = self.agent.train(experience).loss
step = self.agent.train_step_counter.numpy()
if step % hyper_parameters['log_interval'] == 0:
print('step = {0}: loss = {1}'.format(step, train_loss))
if step % hyper_parameters['eval_interval'] == 0:
# Agent Policy Output
avg_return = self.compute_avg_return(self.eval_env, self.agent.policy,
hyper_parameters['num_eval_episodes'])
returns.append(avg_return)
print('step = {0}: Average Return = {1}'.format(step, avg_return))
if ('max_train_reward' in hyper_parameters) and \
(avg_return > hyper_parameters['max_train_reward']):
break
print('\n-----\nTraining complete\n-----')
def choose_actions(self, tracks, sensors, timestamp, nchoose=1, **kwargs):
"""Returns a chosen [list of] action(s) from the action set for each sensor.
Chosen action(s) is selected by exploiting the reinforcement learning agent's
policy that was found during training.
Parameters
----------
tracks: set of :class:`~Track`
Set of tracks at given time. Used in reward function.
sensors: :class:`~Sensor`
Sensor(s) used for observation
timestamp: :class:`tf_agents.trajectories.TimeSpec`
Timestep of environment at current time
nchoose : int
Number of actions from the set to choose (default is 1)
Returns
-------
: dict
The pairs of :class:`~.Sensor`: [:class:`~.Action`] selected
"""
configs = [dict() for _ in range(nchoose)]
for sensor_action_assignment in configs:
for sensor in sensors:
chosen_actions = []
action_step = self.agent.policy.action(timestamp)
action = action_step.action
stonesoup_action = self.env.generate_action(action, tracks, sensor)
chosen_actions.append(stonesoup_action)
sensor_action_assignment[sensor] = chosen_actions
return configs
Create Sensor Managers
We initiate our reinforcement learning sensor manager with the environment we have designed
from stonesoup.sensormanager import BruteForceSensorManager
reinforcementsensormanager = ReinforcementSensorManager({sensorA}, env=StoneSoupEnv())
bruteforcesensormanager = BruteForceSensorManager({sensorB}, reward_function=reward_function)
Train RL agent
To generate a policy, we need to train the reinforcement learning agent using the environment we created above. Some hyperparameters are created that the agent uses to train with.
To train the agent, the hyperparameters are passed to the train method in the ReinforcementSensorManager.
import tensorflow as tf
import reverb
from tf_agents.agents.dqn import dqn_agent
from tf_agents.drivers import py_driver
from tf_agents.networks import sequential
from tf_agents.policies import py_tf_eager_policy, random_tf_policy
from tf_agents.replay_buffers import reverb_replay_buffer, reverb_utils
from tf_agents.specs import tensor_spec
from tf_agents.utils import common
num_iterations = 10000
initial_collect_steps = 100
collect_steps_per_iteration = 1
replay_buffer_max_length = 100000
batch_size = 64
learning_rate = 1e-4
log_interval = 500
num_eval_episodes = 10
eval_interval = 1000
fc_layer_params = (100, 50)
# ---- Optional ----
max_train_reward = 250
hyper_parameters = {'num_iterations': num_iterations,
'initial_collect_steps': initial_collect_steps,
'collect_steps_per_iteration': collect_steps_per_iteration,
'replay_buffer_max_length': replay_buffer_max_length,
'batch_size': batch_size,
'learning_rate': learning_rate,
'log_interval': log_interval,
'num_eval_episodes': num_eval_episodes,
'eval_interval': eval_interval,
'fc_layer_params': fc_layer_params,
'max_train_reward': max_train_reward}
reinforcementsensormanager.train(hyper_parameters)
WARNING:tensorflow:From /home/sgboakes/venvs/stonesoup-rl/lib/python3.10/site-packages/tensorflow/python/util/dispatch.py:1176: calling foldr_v2 (from tensorflow.python.ops.functional_ops) with back_prop=False is deprecated and will be removed in a future version.
Instructions for updating:
back_prop=False is deprecated. Consider using tf.stop_gradient instead.
Instead of:
results = tf.foldr(fn, elems, back_prop=False)
Use:
results = tf.nest.map_structure(tf.stop_gradient, tf.foldr(fn, elems))
step = 500: loss = 1354.437255859375
step = 1000: loss = 19099.845703125
step = 1000: Average Return = 65.86003875732422
step = 1500: loss = 7924.47900390625
step = 2000: loss = 44789.375
step = 2000: Average Return = 117.9991226196289
step = 2500: loss = 94972.765625
step = 3000: loss = 14028.505859375
step = 3000: Average Return = 65.43219757080078
step = 3500: loss = 30156.640625
step = 4000: loss = 4416.0810546875
step = 4000: Average Return = 106.593994140625
step = 4500: loss = 30297.248046875
step = 5000: loss = 6728.6787109375
step = 5000: Average Return = 67.36097717285156
step = 5500: loss = 2970.46826171875
step = 6000: loss = 4445.5810546875
step = 6000: Average Return = 99.21797180175781
step = 6500: loss = 2844.02392578125
step = 7000: loss = 20083.6875
step = 7000: Average Return = 574.1416625976562
-----
Training complete
-----
Run the sensor managers
The choose_actions() function requires a time step and a tracks list as inputs.
For both sensor management methods, the chosen actions are added to the sensor and measurements made. Tracks which have been observed by the sensor are updated and those that haven’t are predicted forward. These states are appended to the tracks list.
Run reinforcement learning sensor manager
To be able to exploit the policy generated by the reinforcement sensor manager, it must be passed appropriate ‘timesteps’. These are distinct from the timesteps in Stone Soup, and is of the form time_step_spec from tf-agents.
from itertools import chain
sensor_history_A = dict()
timesteps = []
for state in truths[0]:
timesteps.append(state.timestamp)
tf_timestep = reinforcementsensormanager.test_env.reset()
reinforcementsensormanager.env.reset()
for timestep in timesteps[1:]:
# Generate chosen configuration
# i.e. {a}k
# Need to make our own "timestamp" that matches tensorflow time_step_spec
observation = []
uncertainty = []
for target in tracksA:
x_target = target.state.state_vector[0] - sensorA.position[0]
y_target = target.state.state_vector[2] - sensorA.position[1]
bearing_target = mod_bearing(np.arctan2(y_target, x_target))
uncertainty.append(np.trace(target.covar))
# observation.append(np.degrees(bearing_target))
observation.append(np.trace(target.covar))
observation = np.array(uncertainty, dtype=np.float32)
# observation = np.array(observation, dtype=np.float32)
chosen_actions = reinforcementsensormanager.choose_actions(tracksA, [sensorA], tf_timestep)
# Create empty dictionary for measurements
measurementsA = []
for chosen_action in chosen_actions:
# chosen_action is a pair of {sensor, action}
for sensor, actions in chosen_action.items():
sensor.add_actions(list(chain.from_iterable(actions)))
sensorA.act(timestep)
# Store sensor history for plotting
sensor_history_A[timestep] = copy.copy(sensorA)
# Observe this ground truth
# i.e. {z}k
measurements = sensorA.measure(OrderedSet(truth[timestep] for truth in truths), noise=True)
measurementsA.extend(measurements)
hypotheses = data_associator.associate(tracksA,
measurementsA,
timestep)
for track in tracksA:
hypothesis = hypotheses[track]
if hypothesis.measurement:
post = updater.update(hypothesis)
track.append(post)
else: # When data associator says no detections are good enough, we'll keep the prediction
track.append(hypothesis.prediction)
# Propagate environment
action_step = reinforcementsensormanager.agent.policy.action(tf_timestep)
tf_timestep = reinforcementsensormanager.test_env.step(action_step.action)
Plot ground truths, tracks and uncertainty ellipses for each target.
import plotly.graph_objects as go
from stonesoup.functions import pol2cart
plotterA = AnimatedPlotterly(timesteps, tail_length=1, sim_duration=10)
plotterA.plot_sensors(sensorA)
plotterA.plot_ground_truths(truths, [0, 2])
plotterA.plot_tracks(tracksA, [0, 2], uncertainty=True, plot_history=False)
def plot_sensor_fov(fig, sensor_history):
# Plot sensor field of view
trace_base = len(fig.data)
fig.add_trace(go.Scatter(mode='lines',
line=go.scatter.Line(color='black',
dash='dash')))
for frame in fig.frames:
traces_ = list(frame.traces)
data_ = list(frame.data)
x = [0, 0]
y = [0, 0]
timestring = frame.name
timestamp = datetime.strptime(timestring, "%Y-%m-%d %H:%M:%S")
if timestamp in sensor_history:
sensor = sensor_history[timestamp]
for i, fov_side in enumerate((-1, 1)):
range_ = min(getattr(sensor, 'max_range', np.inf), 100)
x[i], y[i] = pol2cart(range_,
sensor.dwell_centre[0, 0]
+ sensor.fov_angle / 2 * fov_side) \
+ sensor.position[[0, 1], 0]
else:
continue
data_.append(go.Scatter(x=[x[0], sensor.position[0], x[1]],
y=[y[0], sensor.position[1], y[1]],
mode="lines",
line=go.scatter.Line(color='black',
dash='dash'),
showlegend=False))
traces_.append(trace_base)
frame.traces = traces_
frame.data = data_
plot_sensor_fov(plotterA.fig, sensor_history_A)
plotterA.fig
Run brute force sensor manager
sensor_history_B = dict()
for timestep in timesteps[1:]:
# Generate chosen configuration
# i.e. {a}k
chosen_actions = bruteforcesensormanager.choose_actions(tracksB, timestep)
# Create empty dictionary for measurements
measurementsB = set()
for chosen_action in chosen_actions:
for sensor, actions in chosen_action.items():
sensor.add_actions(actions)
sensorB.act(timestep)
# Store sensor history for plotting
sensor_history_B[timestep] = copy.copy(sensorB)
# Observe this ground truth
# i.e. {z}k
measurementsB |= sensorB.measure(OrderedSet(truth[timestep] for truth in truths), noise=True)
hypotheses = data_associator.associate(tracksB,
measurementsB,
timestep)
for track in tracksB:
hypothesis = hypotheses[track]
if hypothesis.measurement:
post = updater.update(hypothesis)
track.append(post)
else: # When data associator says no detections are good enough, we'll keep the prediction
track.append(hypothesis.prediction)
Plot ground truths, tracks and uncertainty ellipses for each target.
plotterB = AnimatedPlotterly(timesteps, tail_length=1, sim_duration=10)
plotterB.plot_sensors(sensorB)
plotterB.plot_ground_truths(truths, [0, 2])
plotterB.plot_tracks(tracksB, [0, 2], uncertainty=True, plot_history=False)
plot_sensor_fov(plotterB.fig, sensor_history_B)
plotterB.fig
With a properly trained policy, the ReinforcementSensorManager performs almost as well as the
BruteForceSensorManager. Also, once the policy has been learnt, the time taken to run the
tracking loop is far smaller for the ReinforcementSensorManager than for the
BruteForceSensorManager, which must re-calculate the best actions each time it is run.
Metrics
Metrics can be used to compare how well different sensor management techniques are working. Full explanations of the OSPA and SIAP metrics can be found in the Metrics Example.
from stonesoup.metricgenerator.ospametric import OSPAMetric
ospa_generatorA = OSPAMetric(c=40, p=1,
generator_name='ReinforcementSensorManager',
tracks_key='tracksA',
truths_key='truths')
ospa_generatorB = OSPAMetric(c=40, p=1,
generator_name='BruteForceSensorManager',
tracks_key='tracksB',
truths_key='truths')
from stonesoup.metricgenerator.tracktotruthmetrics import SIAPMetrics
from stonesoup.measures import Euclidean
siap_generatorA = SIAPMetrics(position_measure=Euclidean((0, 2)),
velocity_measure=Euclidean((1, 3)),
generator_name='ReinforcementSensorManager',
tracks_key='tracksA',
truths_key='truths')
siap_generatorB = SIAPMetrics(position_measure=Euclidean((0, 2)),
velocity_measure=Euclidean((1, 3)),
generator_name='BruteForceSensorManager',
tracks_key='tracksB',
truths_key='truths')
from stonesoup.dataassociator.tracktotrack import TrackToTruth
associator = TrackToTruth(association_threshold=30)
from stonesoup.metricgenerator.uncertaintymetric import SumofCovarianceNormsMetric
uncertainty_generatorA = SumofCovarianceNormsMetric(generator_name='ReinforcementSensorManager',
tracks_key='tracksA')
uncertainty_generatorB = SumofCovarianceNormsMetric(generator_name='BruteForceSensorManager',
tracks_key='tracksB')
Generate a metrics manager.
from stonesoup.metricgenerator.manager import MultiManager
metric_manager = MultiManager([ospa_generatorA,
ospa_generatorB,
siap_generatorA,
siap_generatorB,
uncertainty_generatorA,
uncertainty_generatorB],
associator=associator)
For each time step, data is added to the metric manager on truths and tracks. The metrics themselves can then be generated from the metric manager.
metric_manager.add_data({'truths': truths, 'tracksA': tracksA, 'tracksB': tracksB})
metrics = metric_manager.generate_metrics()
OSPA metric
First we look at the OSPA metric. This is plotted over time for each sensor manager method.
from stonesoup.plotter import MetricPlotter
fig = MetricPlotter()
fig.plot_metrics(metrics, metric_names=['OSPA distances'])
The BruteForceSensorManager generally results in a smaller OSPA distance
than the observations of the ReinforcementSensorManager, reflecting the better tracking performance
seen in the tracking plots. At some times, the OSPA distance for the ReinforcementSensorManager is slightly
lower than for the BruteForceSensorManager. While it is intuitive to think that the brute force algorithm
would always perform better, the brute force algorithm will pick the target that is most uncertain, and the
reinforcement algorithm may pick another target that happens to reduce OSPA distance more.
SIAP metrics
Next we look at SIAP metrics. We are only interested in the positional accuracy (PA) and velocity accuracy (VA). These metrics can be plotted to show how they change over time.
fig2 = MetricPlotter()
fig2.plot_metrics(metrics, metric_names=['SIAP Position Accuracy at times',
'SIAP Velocity Accuracy at times'])
Similar to the OSPA distances the BruteForceSensorManager
generally results in both a better positional accuracy and velocity accuracy than the observations
of the ReinforcementSensorManager.
Uncertainty metric
Finally we look at the uncertainty metric which computes the sum of covariance matrix norms of each state at each time step. This is plotted over time for each sensor manager method.
fig3 = MetricPlotter()
fig3.plot_metrics(metrics, metric_names=['Sum of Covariance Norms Metric'])
This metric shows that the uncertainty in the tracks generated by the ReinforcementSensorManager is a
little higher than for those generated by the BruteForceSensorManager. This is also reflected by the
uncertainty ellipses in the initial plots of tracks and truths.
References
Total running time of the script: ( 3 minutes 27.071 seconds)