Transformer Surrogate Genetic Programming for Dynamic Container Port Truck Dispatching

Abstract

In the wake of burgeoning demands on port logistics, optimizing the operational efficiency of container ports has become a compelling necessity. A critical facet of this efficiency lies in practical truck dispatching systems. Although effective, traditional Genetic Programming (GP) techniques suffer from computational inefficiencies, particularly during the fitness evaluation stage. This inefficiency arises from the need to simulate each new individual in the population, a process that neither fully leverages the computational resources nor utilizes the acquired knowledge about the evolving GP structures and their corresponding fitness values. This paper introduces a novel Transformer-Surrogate Genetic Programming (TSGP) approach to address these limitations. The methodology harnesses the accumulated knowledge during fitness calculations to train a transformer model as a surrogate evaluator. This surrogate model obviates the need for individual simulations, thereby substantially reducing the algorithmic training time. Furthermore, the trained transformer model can be repurposed to generate superior initial populations for GPs, leading to enhanced performance. Our approach synergizes the computational advantages of transformer models with the search capabilities of GPs, presenting a significant advance in the quest for optimized truck dispatching in dynamic container port settings. This work improves the efficiency of Genetic Programming and opens new avenues for leveraging GP in scenarios with substantial computational constraints.