In this code, which compliments our KANQAS, we provide the code that opens up the possibility of harnessing Kolmogorov-Arnold Network for Quantum Architecture Search i.e., namely KANQAS

The code was used on Ubuntu GNU/Linux 22.04.4 LTS (64-bit).

For this project, we use Anaconda which can be downloaded from https://www.anaconda.com/products/individual.

Before proceeding further kindly install and activate the environment using the following command:

conda env create -f kanqas.yml
conda activate kanqas 

It is recommended to make your own environment using (please check managing envioenments for details):

conda create --name <name-of-your-environment>

then

conda activate <name-of-your-environment>

and after install pip

conda install pip

install the following few dependencies listed below:

pip install numpy
pip install torch
pip install qiskit-aer
pip install pykan

please note that pykan module itself inherently requires some specific softwares to run. Hence, after pykan you install the following:

pip install scikit-learn
pip install pyyaml
pip install matplotlib
pip install tqdm
pip install pandas

Phew, you are done now! You are ready to go! So explore and exploit the possibilities with KANQAS!

MLP

python main.py --seed 1 --config 2q_bell_state_seed1 --experiment_name "DDQN/"

and KAN

python main.py --seed 1 --config 2q_bell_state_seed1 --experiment_name "KAQN/"

The configuration for experiments to Bell and GHZ state constructions are in configuration_files/ folder, where the DDQN folder contains Double Deep Q-Network with Multi-Layer Perceptron and KAQN is Double Deep Q-Learning with Kolmogorov Arnold Network.

The results are saved in the results/ folder.

The KAN part of the code is built using the awesome repository pykan! Also it is recommended to check the Author's note in the pykan git if you are planning to use for other applications!

The MLP part of the code is built using the RL-VQE code agent and RL-VQSD code agent

As a motivation for future work towards the interpretability of KAN, we illustrate trained KAN in constructing Bell state where we use the [84,2,12] configuration. The Tensor encoded quantum circuit as input to KAN contains 84 entries because the quantum circuit is encoded into $D\times (N\times(N+5))$ dimension tensor, where $D=6$ corresponds to maximum depth. For more details, please check our paper. We can see that not all the neurons actively contribute to the choice of action, defined as Quantum gates as output of KAN.

Due to the huge dimension of the KAN in the previous picture, the activation function is in between the input and the output layers; the activation functions are not visible. Hence in the following illustration, we explicitly show the trend of the activation function of the trained KAN

@misc{kanqas_code,
 author = {Akash Kundu},
 title = {{KANQAS GitHub}},
 year = {2024},
 publisher = {GitHub},
 journal = {GitHub repository},
 howpublished = {\url{https://github.com/Aqasch/KANQAS_code}},
 commit = {}
}