In ULISSES we are developing networked data-driven calibration algorithms for autonomous CO2 sensor calibration. In a network each sensor has an own belief on the measurement. Sensors that have been recently calibrated using a reference might have a better calibrated sensor model than other sensors that have not been calibrated for long time and their model has become outdated. The idea of networked calibration approaches is to fuse the beliefs of individual sensors and use this fused belief to update the own sensor model.

In this project we have developed two approaches. In the first approach we fuse beliefs of sensors that meet once and therefore had no chance to establish a joint statistic. Accordingly, the belief fusion strategy is based on the Dempster-Shafer theory. A direct application of this approach would be very sensitive to outliers so that we extended the approach to weight sensor beliefs so that outliers are automatically suppressed.

In case of distributed sensors that repeatedly meet, sensors can establish a joint statistic. This allows the use of reinforcement learning techniques to assess the quality of the individual sensor models and identify which sensor model should be trusted most and therefore serve as a reference for the others. To manage the computational complexity we developed deep reinforcement learning algorithms that aims to minimize the uncertainty given the past data and predicted future. The numerical experiments show that good long-term stability can be achieved.