The results provide concrete tools for applications in climate research, energy networks and autonomous systems, among others.
Dealing intelligently with uncertain information
In many sectors - from weather forecasting to technical systems - predictions are based on mathematical models. Those models describe how a system develops, but in reality are never perfect. At the same time, sensor data are often noisy, incomplete or irregularly available.
Abedini focused on how to still make reliable predictions under such conditions. This process, known as data assimilation, revolves around combining theory and measurements into one most accurate picture of reality.
She makes the comparison to a LEGO set whose instructions are partly missing: you cannot rely on the instructions alone, but must also look at the available pieces and other examples to arrive at the correct construction.
Reliable predictions under realistic conditions
The main finding of the research is that predictions can be surprisingly robust even when data is flawed. Abedini shows that systems can continue to track reality well even when data is noisy or partially missing. It also shows that combining different data sources can improve accuracy, if done correctly. A crucial role here is played by the frequency with which data sources are switched: this greatly determines how stable and reliable the predictions are. It is also striking that it appears to be possible to estimate in advance how accurate a prediction will be.
Direct impact on technology and society
The findings have broad societal relevance. In many modern applications, decisions must be made based on incomplete information, such as in autonomous vehicles that continuously analyze their environment, energy networks that balance supply and demand, wireless sensor networks in smart cities, and climate and environmental monitoring systems.
Abedini's research shows that the way sensors are deployed - and how often they switch between different sources of information - directly affects the quality of predictions. This is especially important in systems where data is not continuously available, for example, due to limitations in battery capacity or bandwidth.
Guidelines for more efficient systems
For engineers, the results offer concrete guidelines. For example, it appears that too little switching between data sources can reduce accuracy, while a well-chosen switching frequency actually makes for more stable and better predictions.
In practical applications, such as environmental monitoring with battery-powered sensors, this can be used to determine how often sensors should be active to both save energy and provide reliable information.
The insights are directly applicable in algorithms for robotics, smart infrastructure and climate models. In doing so, they contribute to more efficient and robust systems - an important step in a society that is increasingly dependent on data while resources remain limited.
Nazanin Abedini defends her dissertation May 7 at Vrije Universiteit Amsterdam.