Transport times in container terminals have been widely studied, often with a focus on autonomous vehicles, while manual operations remain prevalent in many ports. Predicting travel times for straddle carriers in container terminals is a challenging problem, impacting the efficiency and productivity of manual container handling.

In our work, we propose a machine learning-based method for predicting travel times, leveraging a unique dataset obtained from a digital twin. This dataset incorporates positional and operational data gathered from IoT devices. The dataset is sourced from the Port of Hamburg, which is one of the 20 largest container ports in the world.

Our method demonstrates better accuracy and performance compared to conventional methods, tailored to the needs of manual container handling. The conducted research provides insights for optimizing container handling in container terminals. It also has implications for the container terminal simulation and an underlying digital twin.

The Capacitated Vehicle Routing Problem with Time Windows (CVRPTW) is a well-known combinatorial optimization problem that extends the classical Vehicle Routing Problem to account for additional real-world constraints such as truck capacity and customer time windows. Recent studies have explored the use of deep graph convolutional networks (GCNs) to predict the arcs that are part of the optimal tour for the Travelling Salesman Problem and related routing problems.

In our talk, we propose a novel context complemented graph convolutional network (CCGCN) which is integrated into a Monte Carlo Tree Search (MCTS) to solve the CVRPTW sequentially. The CCGCN consists of a deep convolutional part that builds efficient CVRPTW graph representations and a context part. The context part processes information of partially built tours to output probabilities on which vertex to add next to the tour, which is used during the expansion of the search tree. For simulating the final solution value based on a given partial solution in the Monte Carlo search tree, we apply a beam search that uses the context complemented graph convolutional network in an autoregressive form to build valid complete solutions for evaluating a given node in the MCTS. Moreover, a variation of the Upper Confidence Bound applied to trees is used in combination with the prediction probabilities given by the network to manage the tradeoff between exploration and exploitation. Evaluations of the proposed heuristic are performed on benchmark instances of the CVRPTW with up to 100 customers; these show promising results.

Online tour planning in logistics involves dynamically incorporating new transportation requests into precomputed tour plans, without foresight into future demands. Prominent examples include medical logistics involving the timely delivery of medical equipment, pharmaceuticals, or even patients. Online decision-making is quite well understood from an algorithmic and worst-case perspective and tight performance bounds are known.

However, the assumption of not having any prior knowledge about future requests seems overly pessimistic. Given the success of machine-learning methods and the available data in many tour planning problems, one may expect to have access to predictions about future requests. However, simply trusting them might lead to very poor solutions as these predictions come with no quality guarantee. In this talk we present recent developments in the young line of research that integrates such error-prone predictions into algorithm design to break through worst case barriers. We discuss algorithmic challenges with a focus on online routing present algorithms with error-dependent performance guarantees and we shortly discuss the choice of error metrics.

This is an overview talk that builds mainly on joint work with G. Bernardini, A. Marchetti-Spaccamela, L. Stougie, and M. Sweering, published at NeurIPS 2022.