Our Services

Reinforcement Learning

Controlling a Nanoscale Visualisation Apparatus

With a control space spanning more than five dimensions, producing a good image is a hard, time-consuming task even for experts. We contributed to the development of a highly flexible simulation of the imaging environment, allowing the problem to be tackled via reinforcement learning. We developed highly parametrizable environments, allowing configurations to support, in particular, imitation learning and curriculum learning in order to automate this challenging control task.

Probabilistic Models

Mapping Road Lanes from Noisy Sensor Data via Graph-Based SLAM

Based on GPS data, relative motion data and visual observations of landmarks and lane boundaries from multiple vehicle traversals, we reconstructed the most likely lane configuration and geometry for segments of highways undergoing roadworks. This involved probabilistic modelling of the observations, multiple stages of SLAM and, importantly, solving the data association problem via a specialised beam search algorithm.

This work resulted in a publication at the Intelligent Vehicles Symposium.

Probabilistic Map Matching of Vehicle Traces

We devised a custom map matching algorithm based on linear-chain Markov random fields which could soundly integrate numerous matching criteria and which, importantly, always took the “off-road hypothesis” into account. This ensured that data points which were matched to road segments could reliably be associated with the respective segment, allowing traversal data to be leveraged for learning tasks without fearing data contamination.

Our algorithm provided a solid foundation for a wide variety of downstream learning problems that relied on accurate road traversal data.

Supervised Learning

Predicting the (Personalised) Future Energy Requirements of Electric Vehicles

Given information on past trips of vehicles, we learnt specialised energy consumption model for every vehicle. Based on deep learning techniques, our model relates past energy consumption to the situation-specific circumstances (type of road, speed limit, temperature, etc.), allowing the energy requirements of future trips to be predicted much more accurately. Our model vastly outperformed alternative approaches based on physical modelling as well as the industry-standard exponential smoothing methods, reducing prediction errors by 50% on average, thus improving customer confidence in electric range predictions.

Predicting the Future Demand for On-Demand Mobility Services

Based on historic demand data spanning several years, we developed an accurate prediction model for both the local demand that is likely to occur in a particular neighbourhood and the total demand for an entire service area. The local areas were determined based on a custom agglomerative clustering technique that ensured homogeneous usage patterns in each area. The best demand prediction models turned out to be custom neural network-based time series prediction models that appropriately accounted for periodicity.

Combinatorial Optimization & Search

Routing with Probabilistic Goals to Incorporate the Search for a Parking Spot

With parking being scarce, especially in inner city locations, a travel route typically does not end when reaching the destination. Assuming known probabilities for being able to park in different places, we devised a specialised, informed routing algorithm that efficiently finds the route that minimises the probabilistic expectation of travel time (including walking times from potential parking spots) while guaranteeing that a parking spot will almost surely be found along the route.

Our algorithm greatly surpassed the state of the art for this type of problem and was patented by the client.

Optimising Collective Service Agent Plans

Given a set of prioritised tasks, the goal was to assign a sequence of tasks to each available agent such that the combined utility of the satisfied tasks is maximised. Each task is associated with a location that the agent is required to travel to, and each agent’s time schedule and location constraints had to be respected. We applied both exact and approximate optimisation techniques, applying constraint programming and stochastic local search methods in order to solve this problem.