Niko/InteractiveRobotLearning
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

initializing bo finished

Browse Source
This commit is contained in:
Niko Feith 2024-02-20 14:41:38 +01:00
parent 5c6750886c
commit d8dd0c4f4b
5 changed files with 141 additions and 103 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
src/interaction_msgs/CMakeLists.txt
View File

@ -10,7 +10,8 @@ find_package(rosidl_default_generators REQUIRED)
rosidl_generate_interfaces(${PROJECT_NAME}
"srv/Query.srv"
"msg/Opt2Rep.msg"
"srv/Task.srv"
"msg/Opt2Task.msg"
"msg/Opt2UI.msg"
"msg/Task2Opt.msg"
"msg/UI2Opt.msg"

0
src/interaction_msgs/msg/Opt2Rep.msg → src/interaction_msgs/msg/Opt2Task.msg
View File

13
src/interaction_msgs/srv/Task.srv Normal file
View File

@ -0,0 +1,13 @@
# Return reward - True: for updating the model, False: for displaying the x_best to the user
bool reward_return
# Number of dimensions of the representation model
uint16 nr_dim
# Number of parameters per dimensions
uint16 nr_parameter
# Next parameters
float32[] x_next
---
# observed parameters
float32[] x_observed
# observed reward
float32 reward

219
src/interaction_optimizers/interaction_optimizers/bayesian_optimization_node.py
View File

@ -4,69 +4,81 @@ from rclpy.node import Node
from rclpy.callback_groups import ReentrantCallbackGroup
from interaction_msgs.srv import Query
from interaction_msgs.msg import Task2Opt
from interaction_msgs.srv import Task
from interaction_msgs.msg import UI2Opt
from interaction_msgs.msg import TaskOrder
from interaction_msgs.msg import Opt2UI
from interaction_msgs.msg import Opt2Rep
from .optimizers.bayesian_optimization import BayesianOptimization
from transitions.extensions.asyncio import AsyncMachine
import asyncio
import numpy as np
import threading
import time
import os
import numpy as np
class BayesianOptimizationNode(Node):
def __init__(self):
def __init__(self, event_loop):
super().__init__('bayesian_optimization_node')
self.event_loop = event_loop
# region Parameters
self.kernel_type = self.declare_parameter('kernel_type', 'Matern').get_parameter_value().string_value
"""
self.acquisition_function_name = (self.declare_parameter('acquisition_function_name', 'EI')
.get_parameter_value().string_value)
self.nr_bo_steps = self.declare_parameter('bo_steps', 100).get_parameter_value().integer_value
self.nr_dimensions = self.declare_parameter('nr_dimensions',1).get_parameter_value().integer_value
self.nr_dimensions = self.declare_parameter('nr_dimensions', 1).get_parameter_value().integer_value
self.nr_policy_parameters = (self.declare_parameter('nr_policy_parameters', 100)
.get_parameter_value().integer_value)
"""
self.kernel_type = self.declare_parameter('kernel_type', 'Matern').get_parameter_value().string_value
self.nr_init = self.declare_parameter('nr_init', 3).get_parameter_value().integer_value
self.cycle_time = (self.declare_parameter('bo_cylce_time', 0.1)).get_parameter_value().double_value
self.task_request_timeout = self.declare_parameter('task_request_timeout', 10).get_parameter_value().integer_value
self.task_request_max_tries = self.declare_parameter('task_request_max_tries', 3).get_parameter_value().integer_value
# endregion
# Subscribers
self.reward_sub = self.create_subscription(Task2Opt, 'interaction/reward', self.reward_callback, 10)
self.ui_sub = self.create_subscription(UI2Opt, 'interaction/ui_response', self.ui_callback, 10)
self.order_sub = self.create_subscription(Task2Opt, 'interaction/task_order', self.task_order_callback, 10)
self.order_sub = self.create_subscription(TaskOrder, 'interaction/order', self.task_order_callback, 10)
# Publishers
self.ui_pub = self.create_publisher(Opt2UI, 'interaction/ui_request', 10)
self.rep_pub = self.create_publisher(Opt2Rep, 'interaction/next_x', 10)
# Service Clients
self.query_client = self.create_client(Query, 'interaction/user_query')
self.task_client = self.create_client(Task, 'interaction/task_srv')
# Bayesian Optimization
self.bayesian_optimization = None
self.current_episodes = 0
self.nr_bo_steps = 0
self.nr_dimensions = 0
self.nr_policy_parameters = 0
self.acquisition_function_name = ''
self.fixed_dimensions = None
self.seed = None
self.lower_bound = None
self.upper_bound = None
# State Machine
states = ['idle',
'initialize_bo',
'user_query',
'processing',
'waiting_for_user_response',
'waiting_for_task_response']
states = [
{'name': 'idle'},
{'name': 'initialize_bo', 'on_enter': 'initialize_bo_fun'},
{'name': 'user_query', 'on_enter': 'user_query_fun'},
{'name': 'sample_x_next', 'on_enter': 'sample_x_next_fun'},
{'name': 'waiting_for_user_response', 'on_enter': 'waiting_for_user_response_fun'},
{'name': 'processing', 'on_enter': 'processing_fun'},
{'name': 'waiting_for_task_response', 'on_enter': 'waiting_for_task_response_fun'}
]
transitions = [
{'trigger': 'order_received', 'source': 'idle', 'dest': 'initialize_bo', 'after': 'after_receiving_order'},
{'trigger': 'order_received', 'source': 'idle', 'dest': 'initialize_bo'},
{'trigger': 'initialization_finished', 'source': 'initialize_bo', 'dest': 'user_query'},
{'trigger': 'non_interactive', 'source': 'user_query', 'dest': 'sampling_x_next'},
{'trigger': 'sampling_finished', 'source': 'sampling_x_next', 'dest': 'processing'},
@ -84,7 +96,7 @@ class BayesianOptimizationNode(Node):
order_sub - receives order
reset_bo with the order parameters
trigger: order_received()
for i=0:nr_episodes
B) initializing_bo -> initialization_finished() -> user_query:
for j = 0:nr_init
@ -93,7 +105,7 @@ class BayesianOptimizationNode(Node):
reward received from task node
fit model
trigger: initialization_finished()
C1.1) user_query -> non_interactive() -> sampling_x_next
service call for user_query
if user_query is False:
@ -101,7 +113,7 @@ class BayesianOptimizationNode(Node):
C1.2) sampling_x_next -> sampling_finished() -> processing
sampling x_next with acqusition function
trigger: sampling_finished()
C2.1) user_query -> interactive() -> waiting_for_user
service call for user_query
if user_query is True:
@ -113,11 +125,11 @@ class BayesianOptimizationNode(Node):
(UI sends x_next to task node)
receives x_next from UI
trigger: user_response_received()
D) processing -> processing_finished() -> waiting_for_task_response
sends x_next to task node
trigger: processing_finished()
E) waiting_for_task_response -> task_response_received() -> user_query
if episode < nr_episodes:
get results from reward_sub
@ -128,8 +140,8 @@ class BayesianOptimizationNode(Node):
if episode = nr_episodes:
completion_pub with the completed results
trigger: order_completed()
Additional Transitions:
Abort: * -> abort() -> idle
"""
@ -149,100 +161,111 @@ class BayesianOptimizationNode(Node):
acquisition_function_name=self.acquisition_function_name,
kernel_type=self.kernel_type, **kwargs)
# region Callback functions
def task_order_callback(self, msg):
self.current_episodes = 0
self.nr_bo_steps = msg.nr_bo_steps
self.nr_dimensions = msg.nr_dimensions
self.nr_policy_parameters = msg.nr_policy_parameters
self.seed = msg.seed
self.fixed_dimensions = msg.fixed_dimensions
self.acquisition_function_name = msg.acquisition_function_name
asyncio.run_coroutine_threadsafe(self.order_received(), self.event_loop)
def reward_callback(self, msg):
pass
def ui_callback(self, msg):
pass
def task_order_callback(self, msg):
pass
async def send_task_request_with_retry(self, request):
for attempt in range(self.task_request_max_tries):
self.get_logger().info(f"Attempt {attempt + 1}, send task request with retry!")
async def call_service_async(self, client, request):
"""
Asynchronously call a ROS 2 service and wait for its response.
"""
# Use an asyncio Future to wait for the response
future = asyncio.Future()
def callback(response_future):
# Set the result of the asyncio Future when the ROS 2 service responds
rclpy.spin_once(self, timeout_sec=0) # Make sure the callback is processed
future.set_result(response_future.result())
client.call_async(request).add_done_callback(callback)
# Wait for the response
response = await future
return response
# endregion
# region Trigger Placeholers
async def order_received(self):
# Placeholder for trigger
pass
async def initialization_finished(self):
# Placeholder for trigger
pass
async def non_interactive(self):
# Placeholder for trigger
pass
async def interactive(self):
# Placeholder for trigger
pass
async def user_response_received(self):
# Placeholder for trigger
pass
async def processing_finished(self):
# Placeholder for trigger
pass
async def task_response_received(self):
# Placeholder for trigger
pass
async def order_completed(self):
# Placeholder for trigger
pass
async def abort(self):
# Placeholder for trigger
pass
future = self.task_client.call_async(request)
try:
response = await asyncio.wait_for(future, self.task_request_timeout)
return response
except asyncio.TimeoutError:
self.get_logger().warning(f"Task request timed out, resending...")
self.get_logger().error(f"Max retries reached for task request: {self.task_request_max_tries}")
return None
# endregion
# State Methods
async def after_receiving_order(self):
async def initializing_bo_fun(self):
self.get_logger().info('Initializing bo...')
async def user_query_state(self):
self.reset_bo(fixed_dimensions=self.fixed_dimensions)
request = Task.Request()
request.reward_return = False
request.nr_dim = self.nr_dimensions
request.nr_parameter = self.nr_policy_parameters
for i in range(self.nr_init):
request.x_next = self.bayesian_optimization.sample_random_next_observation()
response = await self.send_task_request_with_retry(request)
if i < self.nr_init:
self.bayesian_optimization.add_observation(response.reward, response.x_observed, fit=False)
else:
self.bayesian_optimization.add_observation(response.reward, response.x_observed)
await self.initialization_finished()
async def user_query_fun(self):
self.get_logger().info('Deciding wheter to ask the user...')
srv_msg = Query()
response = await self.call_service_async(self.query_client, srv_msg)
if response.interaction:
await self.interactive()
future = self.query_client.call_async(srv_msg)
try:
response = await asyncio.wait_for(future, timeout=2)
if response.interaction:
await self.interactive()
else:
await self.non_interactive()
else:
await self.non_interactive()
async def ask_user_state(self):
except asyncio.TimeoutError:
self.get_logger().info("Timeout for Query Service...")
async def sample_x_next_fun(self):
pass
async def waiting_for_user_response_fun(self):
self.get_logger().info('Waiting for user response...')
async def processing_next_episode(self):
async def processing_fun(self):
pass
async def waiting_for_task_response(self):
async def waiting_for_task_response_fun(self):
pass
async def completing_order(self):
pass
async def abort(self):
pass
def run_asyncio_loop():
loop = asyncio.new_event_loop()
asyncio.set_event_loop(loop)
loop.run_forever()
return loop # Ensure you return the loop reference
def main():
loop = asyncio.new_event_loop()
asyncio.set_event_loop(loop)
threading.Thread(target=loop.run_forever, daemon=True).start()
# Initialize ROS and pass the event loop to the node
rclpy.init()
node = BayesianOptimizationNode(loop)
# Now, the ROS 2 node has the correct event loop reference for asyncio operations
rclpy.spin(node)
node.destroy_node()
rclpy.shutdown()
loop.call_soon_threadsafe(loop.stop)
if __name__ == '__main__':
main()

9
src/interaction_optimizers/interaction_optimizers/optimizers/bayesian_optimization.py
View File

@ -117,8 +117,8 @@ class BayesianOptimization:
x_next = self.acquisition_function(self.gauss_process, self.x_observed, seed=self.seed)
return x_next
def add_observation(self, y_new, x_new):
if self.episode == 0:
def add_observation(self, y_new, x_new, fit=True):
if self.x_observed.shape[0] == 1:
self.x_observed[0, :] = x_new
self.y_observed[0] = y_new
self.best_reward[0] = np.max(self.y_observed)
@ -127,8 +127,9 @@ class BayesianOptimization:
self.y_observed = np.vstack((self.y_observed, y_new))
self.best_reward = np.vstack((self.best_reward, np.max(self.y_observed)))
self.gauss_process.fit(self.x_observed, self.y_observed)
self.episode += 1
if fit:
self.gauss_process.fit(self.x_observed, self.y_observed)
self.episode += 1
def get_best_result(self):
y_max = np.max(self.y_observed)