Combining Hybrid Genetic Algorithms and Feedforward Neural Networks for Pallet Loading in Real-World Applications

The “Distributor’s Pallet Packing Problem” in a real industrial scenario is addressed in this paper. The main goal is to develop a two-stage algorithm capable to provide the spatial coordinates of the placed boxes vertices and also the optimal boxes input sequence, while guaranteeing geometric, stability, fragility constraints and a reduced computational time. Due to NP-hard complexity of the problem, a hybrid genetic algorithm coupled with a feedforward neural network is used. In the first stage, the hybrid genetic algorithm is run several times on each order within a large set of packing instances, using a different fitness weight vector at each iteration, and storing the best chromosomes to form a rich solution set. After its generation, the best solution is chosen for each order, optimizing a new global weighted function. The global optimal weight vector is tuned by hand, relying on a graphical user interface that allows to show, in real-time, the best solution as a function of the global weights. The dataset is then created, keeping track of both local and global weight vectors related to the optimal solution. Hence, the dataset is used to train, validate and test the neural network. In the second stage, the trained neural network is used to provide the optimal pair of fitness weight vectors, allowing to run the hybrid genetic algorithm only one time and to select directly the optimal solution in the set. The proposed algorithm has been tested and validated on several packing instances provided by an industrial company.