Welcome to the Reinforcement Learning Class Assignment project repository! This project is designed to provide you with hands-on experience in developing reinforcement learning (RL) agents. You will engage with essential components of an RL system, implementing policies to solve RL environments. This repository serves as your starting point, containing the foundational structure and files necessary for the assignment.
The repository is organized as follows to facilitate your development process:
├── README.md
├── requirement.txt # Lists all the necessary dependencies to set up your development environment. (DO NOT REMOVE PYTEST [*])
└── src
├── env_hiv.py # Defines the simulation environment your agent will interact with. (DO NOT MODIFY [**])
├── evaluate.py # Contains the evaluation function to assess your agent's performance. (DO NOT MODIFY [**])
├── interface.py # Provides the agent interface, outlining the methods your agent must implement. (DO NOT MODIFY [**])
├── train.py # Training script. (YOUR CODE GOES HERE)
└── main.py # Serves as the entry point for automatically evaluating your agent. (DO NOT MODIFY [**])
[*] pytest is necessary to the automated evaluation system, don't remove this dependency. Add however any dependency that is necessary to run your agent.
[**] Modifying these files will break the automated evaluation system and assign a grade of zero : don't do it, really.
Your primary task is to create and integrate a training script named train.py within this framework. This script should encompass the training logic for your agent, adhering to the specifications outlined in the interface.py file. Specifically, you are expected to:
- Develop an
Agentclass that conforms to the protocol defined ininterface.py. This class should encapsulate the functionality for your agent, including the ability to:- Decide on actions based on observations from the environment.
- Save what it has learned to disk.
- Load a previously saved model for further inference or evaluation.
You should run your training locally (on your laptop, using colab, etc. as you prefer), then save your model and push both your code and your saved model to your GitHub Classroom repo.
When you push to your GitHub Classroom repo, GitHub Classroom will automatically trigger a series of actions, one of which involves running the main.py script. The main.py script acts as the entry point for evaluating your RL agent. It is designed to instantiate your agent, interact with the environment, and report the agent's scores which participate in the final grade.
This challenge is inspired by the 2006 paper by Ernst et al. titled "Clinical data based optimal STI strategies for HIV: a reinforcement learning approach". It is based on the 2004 simulation model of Adams et al. published in the "Dynamic Multidrung Therapies for HIV: Optimal and STI Control Approaches" paper.
You don't have to read these papers (although the first might be a great inspiration for your work, while the second might provide better insights as to the physical quantities manipulated). In particular, your agent doesn't have to mimic that of Ernst et al.! Feel free to be creative and develop your own solution (FQI, DQN, evolutionary search, policy gradients: your choice).
You are provided with the simulator class HIVPatient of an HIV infected patient's immune system. This simulator follows the Gymnasium interface and provides the usual functions. Your goal is to design a closed-loop control strategy which performs structured treatment interruption, ie. a control strategy which keeps the patient healthy while avoiding prescribing drugs at every time step.
The HIVPatient class implements a simulator of the patient's immune system through 6 state variables, which are observed every 5 days (one time step):
T1number of healthy type 1 cells (CD4+ T-lymphocytes),T1starnumber of infected type 1 cells,T2number of healthy type 2 cells (macrophages),T2starnumber of infected type 2 cells,Vnumber of free virus particles,Enumber of HIV-specific cytotoxic cells (CD8 T-lymphocytes).
The physician can prescribe two types of drugs:
- Reverse transcriptase inhibitors, which prevent the virus from modifying an infected cell's DNA,
- Protease inhibitors, which prevent the cell replication mechanism.
Giving these drugs systematically is not desirable. First, it prevents the immune system from naturally fighting the infection. Second, it might induce drug resistance by the infection. Third, it causes many pharmaceutical side effects which may lead the patient to abandon treatment.
There are four possible choices for the physician at each time step: prescribe nothing, one drug, or both.
The reward model encourages high values of E and low values of V, while penalizing giving drugs.
The patient's immune system is simulated via a system of deterministic non-linear equations.
By default, the HIVPatient class' constructor instantiates always the same patient (the one whose immune system was clinically identified by Adams et al.). However, calling HIVPatient(domain_randomization=True) draws a random patient uniformly.
Your task is to write, train, and save an RL agent which interacts with the default patient for at most 200 time steps and optimizes a structured treatment interruption strategy.
The final grade, out of nine points, is made of two parts.
First, your agent will be evaluated on the default patient. There are six score thresholds: every time your agent passes a threshold, you gain one point.
Then, your agent will be evaluated on the population of patients. There are three score thresholds: every time your agent passes a threshold, you gain another point.