Choosing the appropriate forecasting model for predictive parameter control.

Area of Science:

• Computational intelligence
• Optimization algorithms
• Machine learning

Background:

• Stochastic optimization algorithms require careful parameterization for effective performance.
• Adaptive parameter control (APC) dynamically adjusts parameters during optimization based on past performance.
• Time series prediction is a recent approach for projecting future parameter performance.

Purpose of the Study:

• To evaluate the suitability of various time series prediction methods for projecting future parameter performance in optimization.
• To assess the impact of prediction-based parameter control on evolutionary algorithms (EAs).

Main Methods:

• Investigated multiple time series prediction techniques for parameter performance forecasting.
• Analyzed the conformity of standard EA parameters (excluding population size) to prediction method assumptions.
• Compared predictive parameter control against state-of-the-art APC methods.

Main Results:

• Most standard EA parameters align with the assumptions of the evaluated prediction methods.
• Linear regression marginally outperformed other prediction methods, though not significantly.
• Predictive parameter control surpassed existing methods when data met prediction assumptions.

Conclusions:

• Predictive parameter control offers an effective strategy for optimizing stochastic algorithms, particularly evolutionary algorithms.
• The efficacy of predictive control is dependent on the adherence of performance data to the assumptions of the chosen prediction method.
• When assumptions are not met, predictive control does not significantly degrade algorithm performance.