The IceCube Neutrino Observatory is able to measure the all-flavor neutrino flux in the energy range between 100 GeV and several PeV. Due to the different features of the neutrino interactions and the geometry of the detector, all high-level analyses require a selection of suitable events as a first step. However, presently, no algorithm exists that gives a generic prediction of an event's underlying interaction type. One possible solution to this is the use of deep neural networks similar to the ones commonly used for 2D image recognition. The classifier that we present here is based on the modern InceptionResNet architecture and includes multi-task learning in order to broaden the field of application and increase the overall accuracy of the result. We provide a detailed discussion of the network's architecture, examine the performance of the classifier for event type classification and explain possible applications in IceCube.

On-demand mobility systems in which passengers use the same vehicle simultaneously are a promising transport mode, yet difficult to control. One of the most relevant challenges relates to the spatial imbalances of the demand, which induce a mismatch between the position of the vehicles and the origins of the emerging requests. Most ridepooling models face this problem through rebalancing methods only, i.e., moving idle vehicles towards areas with high rejections rate, which is done independently from routing and vehicle-to-orders assignments, so that vehicles serving passengers (a large portion of the total fleet) remain unaffected. This paper introduces two types of techniques for anticipatory routing that affect how vehicles are assigned to users and how to route vehicles to serve such users, so that the whole operation of the system is modified to reach more efficient states for future requests. Both techniques do not require any assumption or exogenous knowledge about the future demand, as they depend only on current and recent requests. Firstly, we introduce rewards that reduce the cost of an assignment between a vehicle and a group of passengers if the vehicle gets routed towards a high-demand zone. Secondly, we include a small set of artificial requests, whose request times are in the near future and whose origins are sampled from a probability distribution that mimics observed generation rates. These artificial requests are to be assigned together with the real requests. We propose, formally discuss and experimentally evaluate several formulations for both approaches. We test these techniques in combination with a state-of-the-art trip-vehicle assignment method, using a set of real rides from Manhattan. Introducing rewards can diminish the rejection rate to about nine-tenths of its original value. On the other hand, including future requests can reduce users’ traveling times by about one-fifth, but increasing rejections. Both methods increase the vehicles-hour-traveled by about 10%. Spatial analysis reveals that vehicles are indeed moved towards the most demanded areas, such that the reduction in rejections rate is achieved mostly there.

The advances in the area of autonomous delivery robots combined with customers' desire for fast delivery, bare potential for same-day delivery operations, specifically with small time windows between ordering and delivery. Most sameday deliveries are operated using a single depot and with vehicles' routes planned and fixed when leaving the depot. In this paper, we relax these two assumptions and focus on ondemand grocery delivery using a fleet of autonomous vehicles or robots. The problem features the opportunity to pick up goods at multiple local stores or depots, for example, supermarkets within the city, and allows robots to perform depot returns prior to being empty, if beneficial. This allows for more agile planning and on average shorter distance to the next depot. We propose a novel dynamic method for the same-day delivery problem, where we aim to deliver orders as fast as possible, minimally within the same day. In each time step (every few seconds or minutes) the following is executed: For each order potential pick-up locations are identified and feasible trips, i.e., sequences to pick up goods and deliver orders, are calculated. To assign trips to robots an integer-linear program is solved. We simulate one day of service in a city under different conditions with up to 30 autonomous robots, 30 depots and 10,500 orders. Results underpin the advantages of the proposed method and show its versatility with respect to different situations.

This paper presents a novel approach to route heterogeneous fleets for flash delivery operations. Flash deliveries offer to serve customers' wishes in minutes. We investigate a scenario that allows to pick up orders at multiple depots with a heterogeneous vehicle fleet leveraging different modes of transportation. We propose the Heterogeneous Vehicle Group Assignment (HVGA) method, which, given a problem state, identifies potential pick-up locations, calculates potential trips for all modes of transportation and last chooses from the set of potential trips. Experiments to analyze the proposed method are executed using a fleet featuring two modes of transportation, trucks and drones. We compare to a state-of-the-art method. Results show that HVGA is able to serve more orders while requiring less total traveled distance. Further, the effects of the fleet size and fleet composition between drones and trucks are examined by simulating three hours of a flash delivery operation in the city center of Amsterdam.