Env_data for training will be updated soon.

This repository contains the official implementation of the papers:

Z. Yang, S. Gao, X. Cheng and L. Yang, "Synesthesia of Machines (SoM)-Enhanced ISAC Precoding for Vehicular Networks With Double Dynamics," in IEEE Transactions on Communications, vol. 73, no. 9, pp. 7967-7984, Sept. 2025

Z. Yang, S. Gao and X. Cheng, "Doubly-Dynamic ISAC Precoding for Vehicular Networks: A Constrained Deep Reinforcement Learning (CDRL) Approach," GLOBECOM 2024 - 2024 IEEE Global Communications Conference, Cape Town, South Africa, 2024, pp. 5417-5422

Citation:

@ARTICLE{10918656,
  author={Yang, Zonghui and Gao, Shijian and Cheng, Xiang and Yang, Liuqing},
  journal={IEEE Transactions on Communications}, 
  title={Synesthesia of Machines (SoM)-Enhanced ISAC Precoding for Vehicular Networks With Double Dynamics}, 
  year={2025},
  volume={73},
  number={9},
  pages={7967-7984},
  keywords={Precoding;Wideband;Training;Integrated sensing and communication;Real-time systems;Signal to noise ratio;OFDM;Accuracy;Vehicle dynamics;Time-domain analysis;Synesthesia of machine;integrated sensing and communication;deep reinforcement learning;hybrid precoding;double dynamics},
  doi={10.1109/TCOMM.2025.3549503}}

@INPROCEEDINGS{10901120,
  author={Yang, Zonghui and Gao, Shijian and Cheng, Xiang},
  booktitle={GLOBECOM 2024 - 2024 IEEE Global Communications Conference}, 
  title={Doubly-Dynamic ISAC Precoding for Vehicular Networks: A Constrained Deep Reinforcement Learning (CDRL) Approach}, 
  year={2024},
  volume={},
  number={},
  pages={5417-5422}}

Integrated Sensing and Communication (ISAC) is a key technology for future vehicular networks, enabling simultaneous communication and environmental sensing. However, the dynamic nature of wireless channels and rapidly moving targets pose a doubly-dynamic challenge, making real-time precoding design extremely difficult.

This project proposes a Constrained Deep Reinforcement Learning (CDRL) framework to adaptively design ISAC precoders in doubly-dynamic scenarios, without requiring perfect prior channel or target information. The method combines:

-Primal-Dual Deep Deterministic Policy Gradient (PD-DDPG) for constraint-aware learning

-Wolpertinger architecture for flexible action selection with varying numbers of users

-CNN-based feature extraction from historical channel and position states

The proposed approach achieves better sensing accuracy and communication performance compared to traditional optimization-based methods, while significantly reducing computational complexity.

The system models a Constrained Markov Decision Process (CMDP) where:

State: Historical channel estimates and position spectra

Action: Updates to the analog precoding matrix

Reward: Negative CRLB (sensing accuracy)

Cost: Negative spectral efficiency (communication performance)

The agent is trained using PD-DDPG with a Wolpertinger-based action selector to handle variable user counts and complex constraints.

Clone the repository and install dependencies:

git clone https://github.com/ZHYang-PKU/DRL-ISAC.git
cd DRL-ISAC

Python 3.8+

PyTorch 1.12+

NumPy, SciPy, Matplotlib

(Optional) GPU support with CUDA