Reinforcement Learning enabled optimisation for Logistic Systems

Abstract

In modern logistic systems, many optimisation problems are NP-hard. Despite the recent advancement of computer hardware and algorithms, one is still bounded by the time-accuracy trade-off. Because of this limitation, some scholars nowadays are now looking for solutions from an entirely different perspective. Instead of searching in the available space to find one-off solutions, scholars use machine learning algorithms to learn from previous optimisation experience to predict the result. The proposed research would further explore the possible applications of reinforcement learning and neural networks to real-world logistic problems such as various vehicle routing, warehouse management and metro operation problems. In-depth research on the application side is needed because reinforcement learning methods are very problem-specific; researchers must find the most optimum architecture, input format and hyperparameters before applying.

The proposed algorithms would include customised state-of-the-art Neural Network architectures, namely self-attention mechanisms, LSTM, to fit application instances. The algorithms would intake the objective and constraint information, then output the probability of each route option. The parameters of the neural network would be optimised using various reinforcement learning methods. It is expected that after training with a large amount of data, the neural network could produce high-accuracy solutions instantly. Some mathematical derivations would be conducted to prove the convergence to strengthen the proposed algorithms' theoretical foundation. Besides, since Neural Networks often output typos and errors, post-processing Neural Networks would also be designed to improve the solution quality.