Route Sequence Prediction Through Inverse Reinforcement Learning and Bayesian Optimization

We consider the problem of predicting the route sequence a driver will follow given a set of stops and additional data on the delivered packages. We propose an innovative approach to route sequence prediction by using inverse reinforcement learning and Bayesian optimization. Given a sample of route sequences followed by a set of drivers, we want to learn a parameterized cost function that drivers are trying to minimize. Predicted route sequences are generated by using a rollout algorithm. We define our loss function as the average score over predicted route sequences from training data. As the loss function is non-differentiable, we resort to Bayesian optimization, which allows us to efficiently find good parameter values given a limited computational budget. We illustrate the proposed approach with the use of real data and show that it can reasonably approximate real routes generated by human drivers. This suggests that it may be possible to develop models which can learn local information available to drivers and which is reflected in the real routes they build. In some applications, especially in urban environments, routes planned by a machine learning model may be more feasible than routes planned by an optimization model.