Industrial AI Podcast Industrial AI: Time Series, ick hör dir trapsen
Jun 5, 2024
In this discussion, Prof. Dr. Marco Huber, a time series forecasting expert from Fraunhofer IPA, and Marc Zöller, a Senior Machine Learning Engineer at GFT, delve into the world of AutoML. They highlight the significance of time series data in industrial automation and how their AutoSKTime tool innovatively integrates statistical and machine learning methods. Surprising insights reveal simple models can outperform complex ones in many scenarios. They also tackle challenges in scalability and usability for industry, emphasizing the role of AutoML in accelerating development.
AI Snips
Chapters
Transcript
Episode notes
AutoML Tailored For Time Series
- AutoSKTime combines AutoML with SKTime to specifically target time series forecasting.
- The project closes a research-to-industry gap by adapting AutoML principles to temporal data.
Native Time-Series Methods Beat Tabular Tricks
- AutoSKTime avoids sliding-window tabular conversions and uses time-series-native methods.
- It integrates statistical, ML, and neural approaches and uses last-window multifidelity to speed tuning.
Speed Tuning With Increasing History
- Use multifidelity by starting with recent subsequences and increasing history to speed model evaluation.
- Fit quick low-fidelity models on the tail and raise fidelity to extrapolate performance before full training.