Performance Improvement of a Parsimonious Learning Machine Using Metaheuristic Approaches

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Mining

Background:

• Autonomous learning algorithms process data streams online, prioritizing minimal computational complexity.
• Parsimonious learning machines (PALMs) offer structural simplicity for these applications.
• PALMs require hyperparameter tuning (thresholds, fuzziness) often relying on expert knowledge or costly optimization.

Purpose of the Study:

• To develop an advanced PALM using a novel metaheuristic optimization technique.
• To reduce dependency on expert knowledge and computationally expensive optimization methods for PALM hyperparameter tuning.

Main Methods:

• A multimethod-based optimization technique (MOT) was employed to optimize PALM hyperparameters.
• The performance of the MOT-enhanced PALM was compared against greedy search, local search, genetic algorithm (GA), and particle swarm optimization (PSO).

Main Results:

• The proposed PALM utilizing MOT demonstrated superior performance compared to greedy search, local search, GA, and PSO in most evaluated scenarios.
• MOT effectively identified optimal hyperparameters, leading to improved autonomous learning algorithm performance.

Conclusions:

• The multimethod-based optimization technique (MOT) offers significant advantages over single-operator optimization methods for tuning autonomous learning algorithms.
• MOT enables efficient hyperparameter selection for PALMs, maintaining a compact model architecture and enhancing overall performance.