Introduction
In the shadows of the current LLM-based AI hype of text, image, video, and (recently) world generation, there is an edge of scientific development with just as many inventions changing our very lives.
The forecasting space has brought us the likes of Weather forecasting, Stock market prediction, Epidemic/Pandemic prediction, flight traffic management, Supply Chain Optimization, navigation optimization, Precision Agriculture, Smart Cities, Renewable Resource Optimization, Sports Analytics,... the list goes on.
All of these developments are marked by their impact in their respective area. (Somehow with much less controversy involved, though.) We do not connect many of these technologies into the same category, as they are dependent on their field of application. The underlying principles remain the same, though:
Any activity in which the future relies on uncertain and complex interplays between data can be optimized with forecasting.
In this master's thesis, we will be developing better Forecasters that can be applied to real-life problems.
If you are interested, and/or want to know what "better" is supposed to mean, please reach out and let's talk about it.
email: niklase@its.uio.no
room: 312
Objectives
- Develop Forecasters that further our ability to look into the future
- Reach beyond current literature and tackle issues of generalization, value, and comparability of forecasters
- Work on your choice of novel methodology (see Projects) and apply it to real-life data.
Projects
So far, I have presented this topic in a more general context. However, there are subcategories to dig into, which form the core of my own research.
- Building a Better Forecasting System
Current Forecasting research overwhelmingly just presents another new and most accurate method. We want more than that. We want something that shows real value in application, can be compared to other approaches, and generalizes to different spaces. Let's find out how to build it. - Forecasting with Guarantees and Constraints
Machine learning is a great tool to model complex relationships. It has, however, no inherent understanding of what it is learning, nor can it inherently adhere to any logical rules. We can fix that by applying constraints on our networks, giving us real guarantees. Such a model won't be able to violate traffic laws, discriminate in hiring processes, or give us nonsensical results in physical systems. - Explainable Forecasting
Trust is one of the less talked about aspects of the adoption of AI in real-life applications. Trust has to be earned, and while guarantees build one part of it, understanding of why a system makes a decision forms another. There are a bunch of ways to interpret current neural networks, but what about constructing a model that is more transparent by design? Let's look inside the black box.
We'll be applying these in the Energy Sector: my own application specialty is Solar Forecasts; however, other domains are possible too.
Note that these topics are not mutually exclusive, nor are they set in stone. The topics are also not restricted to a single student, either (groups of people with similar interests are welcome!).
If you already have strong ideas or visions related to forecasting with time series that are not quite overlapping with what is presented here, feel free to discuss them with me as well. I am sure that we can accommodate you.
Background needed
Necessary is only experience with Python programming and understanding of calculus and linear algebra. The required Machine learning knowledge can be learned in TEK5040.
Nice to have otherwise:
1. Any experience in relevant application fields (e.g., Energy, Finance, Weather, etc).
2. Any experience building an ML/DL model.
3. Grasp of basic statistical and time series methodology.