Modular learning models in forecasting natural phenomena.

Area of Science:

• Machine Learning
• Computational Science

Background:

• Modular models, a type of committee machine, use specialized local models for distinct input space regions.
• Traditional algorithms often automate region allocation, limiting domain expert knowledge integration.

Purpose of the Study:

• To present novel approaches for building modular models with domain expert involvement.
• To explore various training set splitting and model combination techniques.
• To introduce new algorithms for model trees, enhancing piecewise linear modular regression.

Main Methods:

• Developing modular models using hard splits and soft combinations (statistical, deterministic, fuzzy committees).
• Integrating domain expert knowledge into model allocation and selection.
• Presenting new algorithms for model trees (piecewise linear modular regression models).

Main Results:

• Modular models achieved higher accuracy compared to traditional global learning models.
• The resulting modular models offer improved transparency in forecasting natural processes.
• Demonstrated effectiveness on benchmark tests and river flow forecasting problems.

Conclusions:

• Modular modeling approaches enhance forecasting accuracy and model interpretability.
• Incorporating domain expertise leads to more effective and transparent predictive models.
• The presented methods, including model trees, offer a powerful alternative for complex natural process forecasting.