Team repo for Team Dogleg, working on the case Test reinforcement learning to automate well trajectories - Equinor Virtual Summer Internship 2020.
Planning the trajectory of a subsurface well is to a large extent a manual process. This is resoruce demanding and the resulting well path is not usually not the optimal path to the desired target location. It is therefore desirable to automate this process with the goal of simplify the well planning process, saving resources and creating more optimal well paths.
This repository proposes software attempting to automate the planning of a well path trajectory using reinforcement learning. An agent is trained in an environment modelling a subsurface environment with hazards (no-go areas) and a set of target balls. It uses a custom made environment compatible with OpenAI Gym. The reinforcement learning is implemented using a set of implementations of reinforcement learning algorithms from stable-baselines.
This repository offers two ways to quickly train, retrain or load a trained model using a variety of the available stable-baslines reinforcement learning algortihms. The solution has been dockerized for easy "shipping around-ability", however it is also possible to simply execute the python code outside of Docker on your host machine.
It is recommended to run using docker, however if you don't have Docker available it is possible to run it normally on your computer.
Install the requried dependencies:
~$ pip install -r requirements.txt
Note that some of the packages are only compatable with Python versions 3.7 or lower. Verify your Python verison with python -v
and downgrade or consider running with Docker if the version is insufficient.
To run the program, navigate to the gym-drill
folder
~$ cd gym-drill
and run the program using:
~$ python main.py <action> <save_name> <algorithm> <timesteps>* <new_save_name>*
An asterisk next to the argument indicates that the argument is not required. There are three required arguments <action>
,<save_name>
and <algorithm>
.
- The
<action>
argument has 3 valid values: "train", "retrain" and "load". These will train a new agent, retrain an existing agent or load and display an existing agent, respectively. - The
<save_name>
argument can take any string that is a valid filename on your system. When running with<action> == <train>
this will be the saved name of the newly trained agent. When running with "retrain" or "load" this will be the name of the existing agent to retrain or load. - The
<algorithm>
argument specifies which algorithm to use when training. This argument is not case sensitive. For valid algorithms to pass as an argument see available algorithms. When retraining and loading this algorithm must match the algorithm of the existing agent. To keep track of what algorithm is used on a particular agent a naming convention indicating this is recommended.
In addition to the three required arguments there are additional optional arguments, <timesteps>
and <new_save_name>
, when running with the train or retrain argument.
-
The
<timesteps>
argument can be used with both train and retrain and specifies the total amount of timesteps the training should last. If not specified a total of 10 000 timesteps will be used. -
The
<new_save_name>
argument can be used with retrain and specifies the name of the newly trained agent. If not specified the newly trained will overwrite the old agent named<save_name>
The arguments are summarized in the table below.
Action | Argument | Required arguments | Optional arguments |
---|---|---|---|
Train new agent | train | <save_name> <algorithm_used> | <timesteps> |
Further train existing agent | retrain | <save_name> <algorithm_used> | <timesteps> <new_save_name> |
Load and display existing agent in the environment | load | <save_name> <algorithm_used> | - |
To further clarify how to properly run the program three examples demonstrating the three possible actions are shown.
- To train a new agent for a total of 600 timesteps using the DQN algorithm and saving the agent as my_trained_agent do:
~$ python main.py train my_trained_agent DQN 600
- To retrain a previously trained agent named my_previously_trained_agent for a total of 600 timesteps using the DQN algorithm and saving the retrained agent as my_new_retrained_agent do:
~$ python main.py retrain my_previously_trained_agent DQN 600 my_new_retrained_agent
- To load and view an existing agent named my_existing_agent which has been trained with the DQN algorithm do:
~$ python main.py load my_previously_trained_agent DQN
Running with Docker ensures that the code can execute on any computer with Docker installed. For user convenience all interactions with the Docker container is done from a PowerShell script. Should detalis regarding the specific Docker commands be of interest the user is reffered to the script itself. It is recommended that you run the script from a PowerShell terminal as administrator. To get started, make sure you are allowed to execute scripts by setting the execution policy to unrestriced. In PowerShell, run
~$ Set-ExecutionPolicy Unrestricted
Running with Docker is recommended to ensure a satisfying runtime environment. If you haven't already, install Docker for Windows.
Building and running the docker container is done by passing different flags as arguments to the ./run.ps1
script.
Build the Docker container using the -build
flag:
~$ ./run.ps1 -build
This will create a Docker container named auto_well_path
.
Having built the container, you can run it be passing the -run
flag. Running the container will in this case be the same as running the program. Therefore the arguments introduced in the running the program locally section must be passed.
~$ ./run.ps1 -run <action> <save_name> <algorithm> <timesteps>* <new_save_name>*
Tensorboard graphs will be running on http://localhost:6006 and if you choose to load trained models, the results will be displayed on http://localhost:5000.
Compared to running the program locally python main.py
has been replaced with ./run.ps1 -run
.
If you are developing and making changes to the source code, you often have to rebuild the Docker container to see the changes in effect. Frequently re-building and re-running the container as two operations to see the results of small changes can be tiresome. You can avoid this, and build and run in one operation by passing the -auto
flag when running the script. In that case the arguments needed when passing the -run
flag is needed.
~$ ./run.ps1 -auto <action> <save_name> <algorithm> <timesteps>* <new_save_name>*
All previously mentioned flags that can be passed to run.ps1
have a corresponding abbreviaton. A complete summary of the different flag options is shown in the table below.
Action | Flag | Abbreviaton |
---|---|---|
Build container | -build | -b |
Run container | -run | -r |
Build and run container | -auto | -a |
Currently we don't have a fancy script for Linux users. To run with linux you would have to manually pass the commands from the run.ps1
script into the terminal.
Build the container using
~$ docker build -t auto_well_path .
Then, run the conatiner using
~$ docker run -dit --mount type=bind,source="$(pwd)",target=/usr/src/app -p 0.0.0.0:6006:6006 -p 5000:5000 --name auto_well_path_running auto_well_path
If wanted, start the tensorboard server displaying the tensorboard graphs in the browser using
~$ docker exec -dit auto_well_path_running tensorboard --logdir /usr/src/app/tensorboard_logs --host 0.0.0.0 --port 6006
Finally, start the python program using
~$ docker exec -it auto_well_path_running python main.py <action> <save_name> <algorithm> <timesteps>* <new_save_name>*
where the arguments are the same as the ones in the running the program locally section.
The following algorithms to train agents are available
-
DQN. For more information see the stable-baselines documentation. To specify the DQN algorithm when running the program pass "DQN" or "dqn" as the
<algorithm>
argument. -
PPO2. For more information see the stable-baselines documentation. To specify the PPO2 algorithm when running the program pass "PPO2" or "ppo2" as the
<algorithm>
argument.NB: Since the results with the PPO2 algorithm during development have been weak, the support for PPO2 is limited. Functionality for dispaying trained PPO2 agents have not been developed.
More algorithms comming soon!
To tweak parameters related to the training in stable-baselines, see the agent_training file together with the stable baselines documentation for the specific algorithm you want to alter.
If it is desirable to perform your own non-stable-baselines training or just customize the training in general, using the custom made OpenAI Gym compatible environment you can use the custom Gym-Drill environment.
Gym-Drill is a custom, OpenAI Gym compatible environment modelling a subsurface reservoar. Its main purpose is to be used in a reinforcement learning context to train an agent to find the best path from a given starting point, to a set of target balls. The environment inflicts a curvature constraint (dogleg severity limitation) on the agents well path. An option to include Hazards (hard constraints that the path cannot intersect) into the subsurface environment also exist.
Using the custom environment requires OpenAI Gym (duh), matplotlib for displaying well paths and numpy for number crunching. If not already installed, run from the terminal:
~$ pip install gym numpy matplotlib
To install the custom gym environment, navigate to the gym-drill
folder
~$ cd gym-drill
and run:
~$ pip install -e .
Thats it. Happy training!
To use the environment both gym
and gym_drill
must be imported to the python file where an instance of the environment will be created. Notice the difference between the gym-drill
and the gym_drill
folder. To create an instance of the environment, use OpenAI Gyms make()
function and pass the environment name drill-v0
as argument
import gym
import gym_drill
env = gym.make("drill-v0")
This will create an instance of the environment with the agent (bit) starting in position (1000,1000,0) with the values for the Inclination and Azimuth angles beeing drawn uniformly with intervals (0,pi/4) and (0,2pi) respectively. See register function for details.
There are several parameters that can be adjusted to change the environment, both the "physical" aspects and the rewards for different actions.
To initialize an instance of the environment with your own specificed parameters do
env = gym.make(env_name,startLocation = STARTLOCATION, bitInitialization = BIT_INITIALIZATION)
where STARTLOCATION
is of type Coordinate and BIT_INITIALIZATION
is a list/tuple on the form [initial_azimuth_angle, initial_azimuth_angular_velocity, initial_azimuth_angular_acceleration,initial_inclination_angle, initial_inclination_angular_velocity, initial_inclination_angular_acceleration
. An example of creating an environment with custom parameters would be:
import gym
import gym_drill
import numpy as np
from gym_drill.envs.Coordinate import Coordinate
STARTLOCATION = Coordinate(0.0,0.0,0.0)
BIT_INITIALIZATION = [0.0,0.0,0.0,0.0,0.0,0.0]
env_name = 'drill-v0'
env = gym.make(env_name,startLocation = STARTLOCATION, bitInitialization = BIT_INITIALIZATION)
In addition to [overwriting the starting conditions] of the agent (bit), there exist options to toggle hazards in the environment, to train with Monte Carlo simulated environments in order ensure the existence of a feasible path and a episode based log that gives you realtime updates regarding the training.
-
Toggle hazards by passing
activate_hazards = True
as a keyword argument to themake()
function. This will enrich the environment with hazards of amount and size as specified in the environment config file. See the Adjust environment parameters section for details. By default this is set toTrue
-
Toggle Monte Carlo simulated training by passing
monte_carlo = True
as a keyword argument to themake()
function. This will ensure that an agent training in the environment always will be exposed to an environment where a feasible path to all targets exist. This is done by first generating a set of random paths and then populate those paths with targets. The details of the Monte Carlo simulation is specified in the environment config file. See the Adjust environment parameters section for details. By default this is set toFalse
-
Toggle loading of Monte Carlo generated environment by passing
load = True
as a keyword arugment to themake()
function. IfTrue
togheter withmonte_carlo
then it will not generate a new set of Monte Carlo simulated environments, but load form an existing set. It is recommended to use when plotting trained agent to avoid having to generate a new set of Monte Carlo environments -
Toggle the episode log by passing
activate_log == True
as a keyword argument to themake()
function. This will write the amount of steps and total reward from each episode to a file named "drill_log.txt". This log will contain the total amount of steps NOTE: Using the log will greatly reduce performance during training. It is recommended that the log is used when tweaking the reward system or during very thorough training. By default this is set toFalse
.
As an example, if you want to turn of hazards and Monte Carlo simulated training, but see behind the scenes magic written in the log, you would do
env = gym.make("drill-v0",activate_hazards = False,monte_carlo = False,activate_log = True)
The environment and an agent exeperience in the environment is described by a set of variables that control physical attributes, movement limitations, rewards, what is contained in the observation space and more. These are all stored in the environment_config file. If you feel like changing aspects of the environment for yourself by tweaking these variables all you have to do is update the values inside this file.
As the environment is OpenAI gym compatible it has all the attributes and functions you would expect to be in an OpenAI gym environement pr the documentation. These include, but are not limited to:
- A
reset()
function which resets the environment to its initial state. Note that even if you are overwriting the starting conditions the Azimuth and Inclination angle will be drawn randomly, to ensure that the training is not beeing overfitted for one particular starting angle. - A
step()
function that accepts anaction
and executes that action in the environment. In accordance with the documentation thestep()
function returns:- An
observation
object, containing the new state of the environment after the action has been executed - A
reward
(float), which is the reward the agent recieved for exectuing the particular action - A
done
signal (boolean) indicating wheter or not the episode is over - An
info
(dictionary) message containing diagnostic information useful for debugging.
- An
The only deviation from the functions describes in the documentation is that the render()
function that most OpenAI gym environment use to display the environment has been replaced with two seperate functions. display_planes()
for displaying the horizontal (xy) and vertical (zy) planes and display_3d_environment()
which displays the path and environment in a 3D plot.
A 2D version of the environment was us as the stepping stone to develop the complete 3D environment. It is contained in the 2D-version folder. Be aware that it is outdated and does contain some bugs
Last updated 31.07.2020