ML-assisted Vehicle Routing Under Uncertainty

There exist two types of customer accounts associated with industrial supply chain: the vendor-managed (VM) accounts and customer-managed (CM) accounts. VM accounts have fixed service frequency (e.g., “every Monday”) and are known at the beginning of the planning horizon. CM accounts randomly place orders throughout the planning horizon and expect next-day delivery. There are two parameters associated with every VM account: the frequency of visit (FOV) and initial service day (ISD). Given the fixed schedule for visiting customers with VM accounts, as well as the current portfolio of orders from CM accounts, the distributor solves a routing problem to designate the routes of the trucks in each day. We aim to find a delivery schedule (FOV and ISD) for all VM accounts, such that the routing problems result into profitable solutions.

We proposed a metaheuristic procedure to determine delivery plans for VM accounts where we use derivative free optimization (DFO) algorithm to search the delivery schedule design space. We first initialize the FOV. Then we solve a simple MIP to obtain ISD such that the cumulative demand exerted by VM accounts is balanced across the planning horizon. To evaluate the generated delivery plan (i.e., FOV, ISD), we developed a simulation engine to: 1. forecast the probability of realization of an order from each and every CM account; 2. solve the capacitated vehicle routing problem (CVRP) for the routes and obtain optimal cost. If the target cost reduction is satisfied, we will return the delivery schedule, if not, we will go back in the loop and generate a new set of FOV and repeat.

Research Summary

• Developed time series forecasting algorithm to learn from historical customer orders and predict new orders.
• Developed an efficient delivery schedule search procedure with DFO solver and parallelized implementation.