Using Machine Learning to Automatically Plan Concrete Delivery Dispatching

摘要

Demand for concrete, regardless of the geographical location, is increasing globally. While the Ready Mixed Concrete (RMC) industry is facing an ever-increasing demand for concrete, and although concrete is mostly mixed automatically by machines, the allocation of resources remains labour intensive and is mostly handled by experts in RMC dispatching rooms. This thesis introduces a new method for the automatic planning of concrete delivery dispatching. This method uses machine learning to characterises an expert decision making process. Seven supervised machine learning versions with their boosted algorithms are selected for this purpose and tested with field data. Random-Forest outperforms the other selected algorithms. An method is developed to check the feasibility of misclassified deliveries and the results show that around 30% of misclassified solutions are achievable. In this thesis the Ready Mixed Concrete Dispatching Problem (RMCDP) with/without time window is mathematically modelled and used for assessing the quality of experts’ decisions as well as the quality of solutions obtained by machine learning algorithms (Random-Forest). As empirically proved in this thesis, obtaining exact solutions for large scale RMCDPs with available computing facilities is intractable and characterized as NP-hard. Two heuristic methods (Robust-GA and Sequential-GA) are developed to find a near optimum solution in the absence of an exact solution. Four RMCDPs are selected for evaluating the available methods. The results show that Sequential-GA converges 10 times faster than Robust-GA with achieved solutions at 30% less cost. Also, although there is around 10% gap between solutions obtained by Sequential-GA and the optimum solution, the computing time of Sequential-GA is 100 times less than when using the optimization approach. Thus, the best solutions for all 4 test RMCDPs are obtained by comparing the exact solutions (IP and MIP) and Robust-GA and Sequential-GA. Comparing the best solution with the experts’ decisions reveals that the decisions made by experts are near optimum with an 90% of optimality. The accuracy of the machine learning approach is slightly better than the experts’ decisions and Sequential-GA, however, the computing time for the machine learning based method is significantly lower than for the available approaches. This supports the potential of using machine learning techniques in practical Ready Mixed Concrete Dispatching Problems.