AlphaZero for Exact Unitary Synthesis

This repo contains the code used to produce the results of Unitary Synthesis with AlphaZero via Dynamic Circuits. It contains two submodules:
- a fork of turbozero : a full JAX implementation of AlphaZero agents, modified to better fit this project.
- quantum_compilation : a full JAX implemenation of quantum circuit simulation, with support of clean ancillae.

Installation

First clone the repo and its submodules

git clone --recurse-submodules https://github.com/xvalcarce/az_eus

Setup using a miniconda environment

conda env create --file environment.yml
conda activate az_eus
python -m pip install -e turbozero
python -m pip install -e quantum_compilation

In order to use jupyter notebooks (and kernel) inside the az_eus conda environment, run

conda install jupyterlab
pip install ipykernel
python -m ipykernel install --user --name az_eus --display-name "Python (az_eus)"

You can now start jupyter using the jupyter lab command, select the Python (az_eus) kernel.

Training an agent

Start jupyter lab using jupyter lab within the conda env az_eus. Open the train_agent.ipynb notebook. Edit quantum_compilation/config.ini to chose your environment parameters, e.g. number of qubits, ancillae, gateset, connectivity. Edit notebook/train.ini to set agent and training loop parameters, e.g. AlphaZero hyperparameters. Run the notebook. Data will be saved in the data directory.

Synthesis using a trained agent

We provide 3 trained agents, please download them here and extract the archive in the data/ directory.

Start jupyter lab using jupyter lab within the conda env az_eus. Open the synthesis.ipynb notebook. Select the agent type in the second cell of the notebook. Run the compile function on the target unitary of your choice.