Serial vs. Parallel Hybrid AI: A Contingency Framework for Architecture Selection

This repository contains the complete implementation for the paper:

"Serial vs. Parallel Hybrid AI: A Contingency Framework for Architecture Selection Based on Task Characteristics"
Target Journal: Applied Intelligence (Springer)

The code compares Semi-symbolic (serial pipeline) and Concurrency (parallel voting) architectures across multiple datasets to derive a contingency framework for architecture selection based on task characteristics: dependency ($\delta$), error tolerance ($\varepsilon$), and ambiguity ($\alpha$).

---

📋 Table of Contents

• Overview
• Directory Structure
• Installation
• Usage
• Datasets
• Output Files
• Extending the Framework
• Results Summary
• Citation
• License
• Contact
• Acknowledgments

---

Overview

Hybrid AI systems that integrate symbolic reasoning with subsymbolic learning have demonstrated superior performance over pure approaches. However, two distinct hybrid architectures exist: Semi-symbolic (serial pipeline) and Concurrency (parallel voting). Current practice lacks systematic guidance for choosing between them.

This code provides:

• Implementation of both architectures
• Experiments on 8 datasets across 5 domains
• Statistical analysis with 20 runs per condition
• Contingency framework validation
• Reproducible results for the paper

---

Directory Structure

• 📂 my_results/
• 📄 run_comparison.py
• 📂 figures/
• 📄 generate_figures.py
• 📂 shared/
  • 📄 datasets
  • 📄 metrics
  • 📄 architectures

---

Installation

1. Clone the repository

git clone https://github.com/jamalalqundus/code-paper-serial-parallel-hybrid-ai.git
cd code-paper-serial-parallel-hybrid-ai

2. Create a virtual environment (recommended)

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install dependencies

pip install -r requirements.txt

Requirements (requirements.txt):

• numpy>=1.21.0
• pandas>=1.3.0
• scikit-learn>=1.0.0
• xgboost>=1.5.0
• torch>=1.9.0
• torchvision>=0.10.0
• matplotlib>=3.4.0
• seaborn>=0.11.0
• tqdm>=4.62.0
• scipy>=1.7.0
• ucimlrepo>=0.0.3

Usage

Run Experiments

cd experiments
python run_comparison.py --datasets all --runs 20 --output ./my_results

Command Line Arguments

Argument	Description	Default
--datasets	Dataset names (comma-separated) or "all"	"all"
--runs	Number of runs per configuration	20
--epochs	Training epochs per run	5
--output	Output directory	"./my_results"
--list	List available datasets and exit	False

Examples

# Run all datasets
python run_comparison.py --datasets all --runs 20 --output ./my_results

# Run specific datasets
python run_comparison.py --datasets heart_disease,nsl_kdd,adult_income --runs 10 --output ./my_results

# List available datasets
python run_comparison.py --list

# Quick test run
python run_comparison.py --datasets mnist --runs 5 --epochs 3 --output ./test_results

Generate Figures

After running experiments, generate the paper figures:

python generate_figures.py --results ./my_results --output ./figures

Datasets

The following datasets are included (from shared/datasets/loader.py):

Dataset	Domain	Task Type	Samples	Features
UCI Heart Disease	Healthcare	Classification	297	13
NSL-KDD	Cybersecurity	Classification	5,000	41
Adult Income	Social Science	Classification	10,000	14
Credit Card Fraud	Finance	Classification	5,412	30
MNIST	Computer Vision	Classification	10,000	784
HighDep_LowAmb	Synthetic	Classification	1,000	12
LowDep_HighAmb	Synthetic	Classification	1,000	15
Medium_Balanced	Synthetic	Classification	1,000	13

Note: The synthetic datasets are designed to validate the contingency framework with controlled characteristics.

---

Output Files

File	Description
comparison_results.csv	Raw results for all runs (320 rows)
comparison_summary.csv	Aggregated statistics per (dataset, architecture)
contingency_analysis.csv	Contingency factors and accuracy differences
accuracy_comparison.png/pdf	Bar chart comparing architectures
latency_comparison.png/pdf	Latency comparison chart
explainability_comparison.png/pdf	Explainability scores chart
confidence_comparison.png/pdf	Confidence calibration chart
contingency_analysis.png/pdf	Accuracy difference vs ambiguity plot

Extending the Framework

Adding a New Dataset

1. Create a new class in shared/datasets/loader.py inheriting from BaseDataset:

class MyNewDataset(BaseDataset):
    def __init__(self):
        super().__init__(
            name='My_Dataset',
            domain='my_domain',
            task_type='classification'  # or 'regression'
        )
    
    def _load_raw(self):
        # Load your data
        X, y = load_my_data()
        return X, y
    
    def get_characteristics(self):
        return {
            'dependency': 3.0,
            'error_tolerance': 3.0,
            'ambiguity': 3.0
        }

2.Add to get_all_datasets() function:

def get_all_datasets():
    return {
        # ... existing datasets ...
        'my_dataset': MyNewDataset(),
    }

Adding a New Architecture

Create a new class in shared/architectures/base.py inheriting from BaseHybridArchitecture.

---

Results Summary

Finding	Result
Concurrency win rate	5/8 datasets (62.5%)
Improvement on high-ambiguity tasks	+0.2% to +0.9% absolute accuracy
Explainability (Concurrency)	3.0 vs 2.4 (5-point scale)
Latency overhead	5-6× slower
Contingency prediction accuracy	75% (6/8 datasets)

---

Citation

If you use this code in your research, please cite (NOTE: NOT YET PUBLISHED!!):

@article{alqundus2026serialparallelhybridai,
  title={Serial vs. Parallel Hybrid AI},
  author={Al Qundus, Jamal},
  journal={Applied Intelligence},
  year={2026},
  publisher={Springer}
}

License

This project is licensed under the MIT License

License

This project is licensed under the MIT License - see the LICENSE file for details.

---

Contact

Jamal Al Qundus
German Jordanian University
Email: jamal.alqundus@gju.edu.jo
GitHub: jamalalqundus

---

Acknowledgments

This research was supported by the Business Intelligence and Data Analytics department at German Jordanian University.