A Policy-Gradient Reinforcement Learning Framework for Informed Initialization of Global Optimizers

Note for Reviewers: This repository contains the official source code and supplementary materials for the manuscript submitted to Applied Soft Computing (Manuscript Number: ASOC-D-26-11483).

📌 Overview

This project implements a hybrid optimization framework that leverages a lightweight Neural Network (NN) trained via a brief policy-gradient reinforcement learning phase to generate informed initial populations for global optimizers like Particle Swarm Optimization (PSO) and Differential Evolution (DE).

By combining targeted exploitation seeds from the NN with randomly sampled exploration seeds, the framework significantly improves early-stage convergence and final solution quality on complex, multimodal fitness landscapes (evaluated at $D=30$ and $D=50$).

⚙️ Prerequisites & Installation

The code is written in Python 3.12.1 To install the required dependencies, run:

git clone https://github.com/barzansaeedpour/Policy-Gradient-Reinforcement-Learning-Optimization.git

cd Policy-Gradient-Reinforcement-Learning-Optimization

pip install -r requirements.txt

🚀 How to Run

To run an optimization benchmark and reproduce the results discussed in the paper, use the following simple command:

Example: Run the proposed algorithm on the benchmarks for 30 runs

python main.py

📊 Sample Results (Experimental Outputs)

Below are two sample visualization outputs generated by our framework on the multimodal Ackley benchmark function ($D=30$). All output figures, convergence logs, and statistical comparisons are automatically saved in the Result/ directory in high-resolution PDF format.

1. Ackley 3D Landscape & Optimization Path

This plot illustrates the multimodal terrain of the Ackley function alongside the informed initial optimization seeds generated via the policy-gradient reinforcement learning phase.

• View Full PDF: Result/Ackley_3D_path.pdf

---

2. Convergence Analysis on Ackley ($D=30$)

Comparison of convergence speeds and global optimum searching capacity between the Proposed Algorithm, Pure_PSO, DE, GWO, NN_Only, and NN_Init_DE across 30 independent runs, as discussed in the manuscript.

• View Full PDF: Result/Ackley_convergence.pdf

📂 Repository Structure

• main.py: Entry point for running optimization experiments and statistical tests.
• Result/: The generated files, reporting the statistics and plots for comparing.
• requirements/: The required packages to install

✉️ Contact

If you have any questions during the review process, please contact the corresponding author via the journal's editorial system or reach out to:

• Barzan Saeedpour - barzansaeedpour@gmail.com
• Eghbal Hosseini - kseghbalhosseini@gmail.com