Rekomendasi Alokasi Kendaraan Logistik Menggunakan Model Hibrida Transformer Autoencoder dan Algoritma Genetika

Abstract

The vehicle allocation process in logistics operations at PT XYZ is still largely rule-based, making decision-making less effective as delivery requests continue to increase. The growing volume of requests increases the complexity of selecting appropriate vehicles and raises the risk of allocation errors, such as vehicle overloading or assigning unsuitable vehicle types. To address this problem, this study proposes a hybrid model that integrates a Transformer Autoencoder with a Genetic Algorithm for vehicle allocation optimization. The Transformer Autoencoder learns normal operational patterns from historical records of successful vehicle allocations in an unsupervised manner and generates anomaly scores that indicate the suitability of vehicle-load combinations. These anomaly scores are then incorporated into a multi-objective optimization process using a Genetic Algorithm, which evaluates candidate vehicles based on anomaly score (60%), distance to the pickup location (20%), and vehicle capacity utilization (20%). Experimental results demonstrate that the Transformer Autoencoder effectively captures normal operational patterns, achieving a reconstruction loss of 0.0214 after training. The Genetic Algorithm further improves allocation quality by producing the best vehicle recommendations with a fitness score of 0.8263 and an anomaly score of 0.0411. The proposed hybrid approach successfully filters unsuitable vehicle candidates before optimization, resulting in more accurate, efficient, and operationally consistent vehicle allocation recommendations for logistics operations.