Genetic Programming with Reinforcement Learning Trained Transformer for Real-World Dynamic Scheduling Problems
Xian Chen, Rong Qu, Jing Dong, Ruibin Bai, Yaochu Jin·April 10, 2025
Summary
GPRT, using Transformer-enhanced Genetic Programming, addresses dynamic scheduling with real-world unpredictability. It adapts heuristics for container terminal truck scheduling, outperforming traditional methods. The framework employs an RNN to generate heuristics, combined with GP outputs, enhancing heuristic generation through training on fitness scores. Replacing RNN with a Transformer improves handling of long-range dependencies, surpassing previous models in dynamic scheduling environments. The paper's application demonstrates superior prediction, generating heuristics within the GPRT framework. The Transformer's self-attention mechanism boosts scheduling efficiency and understanding, making it ideal for dynamic challenges. Real-world container port scheduling experiments highlight its superiority, showing enhanced operational efficiency and robustness.
Introduction
Background
Overview of dynamic scheduling challenges in container terminal truck operations
Importance of real-world adaptability in scheduling algorithms
Objective
To present GPRT, a novel approach that integrates Transformer-enhanced Genetic Programming for dynamic scheduling
Highlight the framework's ability to adapt heuristics for container terminal truck scheduling, surpassing traditional methods
Method
Data Collection
Gathering real-world data for container terminal operations
Importance of accurate data for effective scheduling
Data Preprocessing
Preparing data for model training
Techniques for handling real-world unpredictability
Model Architecture
Description of the GPRT framework
Integration of RNN and GP for heuristic generation
Training process on fitness scores to enhance heuristic quality
Transformer Integration
Replacing RNN with a Transformer for improved handling of long-range dependencies
Explanation of the Transformer's self-attention mechanism and its benefits in dynamic scheduling
Results
Prediction Accuracy
Demonstration of superior prediction capabilities of GPRT
Comparison with traditional scheduling methods
Heuristic Generation
Description of how GPRT generates heuristics within the framework
Analysis of the enhanced heuristic quality
Efficiency and Robustness
Real-world container port scheduling experiments showcasing operational efficiency and robustness
Comparison with previous models in dynamic scheduling environments
Conclusion
Summary of GPRT's Contributions
Recap of GPRT's innovative approach to dynamic scheduling
Highlighting the framework's adaptability and performance improvements
Future Directions
Potential areas for further research and development
Suggestions for integrating GPRT in broader logistics and scheduling contexts
Advanced features