Application of Reinforcement Learning Algorithms for the Adaptive Computation of the Smoothing Parameter for...

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computational Statistics

Background:

• Probabilistic Neural Networks (PNNs) are powerful classification tools.
• Optimizing the smoothing parameter (sigma) is crucial for PNN performance.
• Current methods for PNN smoothing parameter selection have limitations.

Purpose of the Study:

• To propose novel methods for selecting and adapting the smoothing parameter in PNNs.
• To apply reinforcement learning algorithms for PNN optimization.
• To enhance the predictive accuracy of PNN classifiers.

Main Methods:

• Utilized Q(0)-learning, Q(λ)-learning, and stateless Q-learning algorithms.
• Investigated three PNN models: single parameter, attribute-specific parameter, and class-variable matrix parameter.
• Employed cross-validation to evaluate test error on eight diverse databases.

Main Results:

• The proposed reinforcement learning-based methods achieved competitive results compared to state-of-the-art techniques.
• Demonstrated effective smoothing parameter optimization for various PNN architectures.
• The new approaches showed potential as viable alternatives for PNN training.

Conclusions:

• Reinforcement learning offers an effective framework for PNN smoothing parameter optimization.
• The proposed methods provide robust alternatives for training PNN models.
• Further research can explore advanced reinforcement learning strategies for PNNs.