The group method of data handling (GMDH) is a family of inductive methods for computer-based mathematical modelling of multi-parametric datasets with completely automatic structural and parametric model optimization.

In 1968, Kyiv's Institute of Cybernetics professor Alexey G. Ivakhnenko developed the technique. Since the inductive method was initially implemented on a computer, the key practical outcomes at the foundation of the new theoretical concepts were a collection of computer programs and algorithms. 

As more and more mundane tasks are automated, the possibility of subjective bias in the final product decreases. The AI thesis, which claims that a machine can serve as an influential counsellor to people, finds an implementation in this method.

GMDH algorithms

GMDH algorithms are distinguished by an inductive method that sorts out gradually complicated polynomial models and selects the best solution using an external criterion. Using inductive GMDH algorithms allows for automatically discovering data correlations, choosing a model's or network's most efficient architecture, and improving current algorithms' precision.

This novel self-organizing strategy is distinct from the traditional inductive approaches typically utilized in the modelling industry. It's inductive, meaning it discovers the best answer by eliminating the others. By putting different options in order, GMDH networks try to ensure that the author doesn't have too much effect on the modelling results. The computer determines the best way for a model or set of rules to be put together.

Data Handling

Group Method of Data Handling is a group of algorithms that can be used to solve different tasks. It comprises parametric, clusterization, alternatives complexing, binarization, and probability algorithms. This inductive method sorts out more complicated models and chooses the best answer based on the fewest external criteria. 

Benefits:

• The optimal complexity of the model structure is discovered, which is appropriate for the noise level in the data sample. Simplified optimum models are more accurate for real-world problems with noisy or short data.
• The number of hidden layers and neurons, model structure, and other appropriate hyperparameters are all determined automatically.
• It ensures that the most accurate or unbiased models are found - the approach does not overlook the best answer while sorting all variants (in the provided class of functions).
• Non-linear functions or features can be employed as input variables, which may affect the output variable.
• It detects interpretable relationships in data and chooses effective input variables automatically.
• For software development, GMDH sorting algorithms are relatively straightforward.
• Other modelling algorithms can benefit from the use of twice-multilayered neural nets.
• The method obtains information directly from a data sample and reduces the influence of prior author assumptions about modelling results.
• The approach allows for discovering an objective physical model of an item (law or segmentation) - the same on subsequent samples.

Conclusion

When dealing with the challenges of structural-parametric model identification for unknown experimental data, GMDH is the gold standard solution. Constructing a mathematical model representing the researched object or process leads to this issue. It takes advantage of knowledge about it that is hidden in data. The ideas of automatic model generation, inconclusive decisions, and consistent selection by external criteria for discovering models of optimal complexity set GMDH apart from previous modelling approaches. 

Deep learning networks now employ such a method. Two or more subsets of a data sample are used to compare and select optimal models. Since sample division implicitly acknowledges multiple kinds of uncertainty during the automatic creation of the optimal model.

Image source: Unsplash