Lower upper bound estimation method for construction of neural network-based prediction intervals.

Area of Science:

• Machine Learning
• Statistics
• Computational Science

Background:

• Traditional neural network (NN) prediction intervals (PIs) often rely on strict data assumptions and are computationally intensive.
• Quantifying uncertainty in point forecasts is crucial for reliable predictions.

Purpose of the Study:

• To develop a novel, efficient, and dependable method for constructing PIs for NN predictions.
• To address the limitations of existing PI construction techniques in terms of computational cost and distributional assumptions.

Main Methods:

• Proposed the Lower Upper Bound Estimation (LUBE) method, utilizing a two-output NN to estimate PI bounds.
• Developed a PI-based objective function to optimize both interval width and coverage probability during NN training.
• Employed simulated annealing for cost function minimization and NN parameter tuning.

Main Results:

• The LUBE method demonstrated the ability to generate high-quality PIs rapidly across 10 benchmark regression problems.
• Quantitative comparisons showed LUBE to be simpler, faster, and more reliable than three traditional PI construction techniques.
• The method effectively quantifies forecast uncertainty without requiring pre-defined PI bounds for training.

Conclusions:

• The LUBE method offers a significant advancement in constructing accurate and efficient prediction intervals for neural network models.
• This approach provides a more accessible and computationally feasible alternative for uncertainty quantification in forecasting.
• LUBE enhances the reliability and practical applicability of neural network predictions in various regression tasks.