Y1fanHE / kdps-arm

Code and data repository for the paper titled as "Incorporating Sub-programs as Knowledge in Program Synthesis by PushGP and Adaptive Replacement Mutation"

Knowledge-Driven Program Synthesis - Adaptive Replace Mutation

Code and data repository for the paper titled as "Incorporating Sub-programs as Knowledge in Program Synthesis by PushGP and Adaptive Replacement Mutation"

Table of Contents

• General idea of the Knowledge-Driven Program Synthesis
• Getting started
• Usage
• Detailed design of the conceptual system

General idea of the Knowledge-Driven Program Synthesis

Objectives

Genetic programming (GP) searches a program by random initialization, unguided variation, and fitness-guided selection. This shows a difference to a human programmer. Usually, a human programmer writes code based on his knowledge from his previous programming experience.

This study aims to build a system so that GP could use the knowledge from the previous problems it solved in a sequence of problem-solving (shown in the above figure). We use sub-programs as knowledge in the current study.

Click here to see a detailed description of knowledge.

↑back to top

General framework of Knowledge-Driven Program Synthesis (simple version)

Details

The flowchart above shows a simple version of our framework of the Knowledge-Driven Program Synthesis (KDPS) system. It works as follows.

• The system saves knowledge from the program generated by GP to an Offline Knowledge Archive (Offline KA).
• The system creates an Online Knowledge Archive (Online KA) from the Offline KA.
• The system selects knowledge from the Online KA to the GP.
• The system evaluates the knowledge based on the feedback from GP.

A more detailed design of our conceptual system is in Detailed design of the conceptual system.

↑back to top

Framework in the current study

Details

In the current study, we ommit the Offline KA and focus on the implementation of Online KA as the first step. Therefore, the steps related to Offline KA are replaced by Human.

↑back to top

Getting started

Preparation

1. Install Python 3 (< 3.10) on your computer

2. Install required libraries

   pip install numpy pyyaml

↑back to top

Installation

1. Clone the entire repository

   git clone https://github.com/Y1fanHE/pushgp-adaptive-replacement-\
   mutation-with-knowledge.git

2. Install the modified PyshGP with the following commands

   cd pyshgp
   pip install .

3. Download data files of the benchmark problems and expand the archive

   tar zxvf psgb.tgz

   This command should create a folder named psgb including data files of the four problems at the root of this repository.

4. Test your installation

   python 02/umad.py 1000

   The above command runs PushGP with UMAD to solve the Small or Large problem with seed 1000.

↑back to top

Usage

Experiments

First, copy the Shell scripts in shell_script folder to the repository root

cp shell_script/* .

In the paper, we totally performed the following experiments. We list them with the command to run these experiments.

1. Experiment I: adaptive replacement mutation under an ideal condition

   • Prepare the knowledge archives & Effectiveness of the adaptive replacement mutation

     ./exp102.sh && ./exp104.sh && ./exp114.sh && ./exp127.sh

   • Robustness of the adaptive replacement mutation

     ./exp202.sh && ./exp204.sh && ./exp214.sh && ./exp227.sh

   • Tracking the tested-good knowledge

     ./exp302.sh && ./exp304.sh && ./exp314.sh && ./exp327.sh

2. Experiment II: adaptive replacement mutation under a realistic condition

   ./exp402.sh && ./exp404.sh && ./exp414.sh && ./exp427.sh
   ./expc01.sh && ./expc02.sh

↑back to top

Check the results

• The raw data of the experiments is stored in the 02/dat, 04/dat, 14/dat, 27/dat, c01/dat, and c02/dat folders.

• Copy the Jupyter notebooks in the folder notebook to the repository root.

  cp notebook/* .

  Now, you can also check the results with exp1.ipynb, exp2.ipynb, exp3.ipynb, and exp4.ipynb.

• The figures used in the manuscript are generated by the Jupyter Notebooks and move to the fig folder by the following command.

  ./rename_img.sh

↑back to top

Quick view of the results

• The success rate of five methods on four benchmark problems in Experiment I

↑back to top

Detailed design of the conceptual system

This section describes the detailed design of our conceptual system in four parts, namely GP & Online KA, GP & Offline KA, Online KA & Offline KA, and Offline KA.

GP and Online Knowledge Archive

Details

• The Selector selects knowledge from the Online KA. This knowledge could be mutated by the Mutator.
• The Extractor extracts knowledge during the optimization of GP.
• The GP returns the knowledge applied into the search process before.
• The Evaluator evaluates the quality of the knowledge, either extractrd or returned directly from GP, based on the evolutionary process.
• The Online KA collects the evaluated knowledge to update the unevaluated ones

↑back to top

GP and Offline Knowledge Archive

Details

• The Extractor extracts knowledge from the program generated by GP.
• The Offline KA collectes the extracted knowledge.

↑back to top

Online KA and Offline Knowledge Archive

Details

editing ...

↑back to top

Offline Knowledge Archive

Details

editing ...

↑back to top

License

Distributed under the MIT License. See LICENSE for more information.

↑back to top

Contact

• Yifan He ([e-mail])
• Claus Aranha ([e-mail])

↑back to top

Acknowledgements

Our implementation is based on PyshGP. We are grateful to @erp12 for his effort on developing this library.

The data files of the benchmark problems are from an GitHub repository. We express our thanks to @thelmuth for this wonderful sharing.

↑back to top