Achieving Simultaneous Buffering and Steiner Tree Synthesis via Harmonic-Based Reinforcement Learning

Official implementation for the paper "Achieving Simultaneous Buffering and Steiner Tree Synthesis via Harmonic-Based Reinforcement Learning".

Summary

This project extends an RL-based Rectilinear Steiner Tree (RST) framework to jointly synthesize Steiner topology and predict buffer candidates using a harmonic-function parameterization. The code implements an actor–critic model that outputs routing/topology decisions together with buffering information.

Repository layout (key files)

• train.py — training script (uses models2 and utils2)
• test.py — evaluation / test script
• README.md — this file
• models2/
  • actor_critic.py — actor and critic architectures
  • self_attn.py — encoder / attention modules
  • utils.py — model helper utilities
• utils2/
  • rsmt_utils5.py — evaluation utilities (interfaces to Steiner evaluation)
  • log_utils.py — logging helpers
• data/, nets_data/, nets_data/* — example nets and datasets used for evaluation
• logs/ — TensorBoard logs (created by train.py)
• saved_model/sinModel/ — model checkpoints saved by training
• results/ — evaluation outputs

Requirements

• Python 3.8+
• PyTorch
• NumPy
• tqdm
• TensorBoard

Quick start

1. Train (example):

python train.py --degree 10 --batch_size 256 --num_batches 10000 --learning_rate 5e-5

• Models and checkpoints saved under saved_model/sinModel/.
• TensorBoard logs written to logs/sinModel//

2. Evaluate / test:

python test.py --degree 10

3. View training curves:

tensorboard --logdir=logs

Notes

• train.py expects utilities in models2/ and utils2/. Evaluator implemented in utils2/rsmt_utils5.py performs batch evaluation of routing/buffering decisions.
• Saved checkpoints: saved_model/sinModel/.pt and saved_model/sinModel/b.pt
• Random seeds are set in train.py via the --seed argument for reproducibility.

Acknowledgments

The implementation builds on the REST RSMT work: Liu, J., Chen, G., & Young, E. F. (2021). REST: Constructing Rectilinear Steiner Minimum Tree via Reinforcement Learning. DAC 2021.