Developing Artificial Intelligence Agents to Manipulate Quantum Entanglement

This repository provides all the code needed to run the experiments that support the claims in the text of the master thesis titled `Developing Artificial Intelligence Agents to Manipulate Quantum Entanglement'.

How to run the code

Download the repository locally:

git clone https://github.com/cacao-macao/entanglement-control.git

The code was written using Python 3.10. All of the dependencies can be found in the requirements.txt package and installed using:

pip3 install -r requirements.txt

To start a training procedure run the following script from inside the scripts folder:

cd path-to-repo/entanglement-control/scripts
python3 pg_train_agent.py -q 5 --env_batch 1024 --steps 40 -i 10001 --ereg 0.01

The results from the training procedure will be stored inside the logs folder. Note that running 10001 iterations on a 5-qubit quantum system takes around 19 hours on a TeslaT4 GPU.

How to read the code

The code in the repository has to the following structure:

  |--- scripts
  |   |--- ...      # scripts for running different simulations
  |--- src
  |   |--- agents   # implementations of different RL agents
  |   |--- envs     # implementation of a quantum simulator
  |   |--- infrastructure # logging and utilities
  |   |--- policies # implementations of policies using deep learning

The following agents for controlling entanglement are implemented:

• base_agent.py provides an interface for the agent object
• ac_agent.py implements an actor-critic agent following the advantage actor-critic algorithm
• pg_agent.py implements a policy gradient agent following the vanilla policy gradient algorithm
• il_agent.py implements an imitation learning agent following the behaviour cloning algorithm
• expert.py implements an agent using beam search

For implementing the agents the following API is used:

• agents have a policy and an environment object,
• every agent has a rollout() method that starts an agent-environment interaction and produces an episode,
• every agent has a train() method that starts producing rollouts and uses the experiences from those rollouts to update the policy parameters.

The main policy model used in the experiments is the fcnn_policy.py. This package implements a fully-connected neural network using PyTorch.