Improving the performance of mutation-based evolving artificial neural networks with self-adaptive mutations.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Neuroevolution offers a gradient-free method for evolving artificial neural networks (ANNs), simultaneously optimizing topology and weights.
• Traditional neuroevolution with topological evolution faces challenges with crossover due to the competing conventions problem.
• Mutation-based neuroevolution avoids crossover, relying solely on mutations for genetic variation, presenting an alternative approach.

Purpose of the Study:

• To enhance the performance of mutation-based artificial neural network evolution.
• To introduce a self-adaptive mutation mechanism for improved exploration-exploitation balance.
• To mitigate topological bloat by dynamically adjusting structural mutation probabilities based on network size.

Main Methods:

• Implemented a self-adaptive mutation mechanism to automatically adjust mutation step size.
• Developed a method to dynamically adjust structural mutation probabilities according to network size.
• Evaluated the proposed methods on locomotion tasks using OpenAI Gym benchmarks.

Main Results:

• The proposed self-adaptive mutation mechanism significantly improved performance compared to conventional neuroevolution algorithms.
• Adjusting structural mutation probabilities effectively reduced topological bloat while sustaining performance.
• The enhanced mutation-based approach demonstrated superior results in evolving neural network architectures.

Conclusions:

• Self-adaptive mutation mechanisms and dynamic structural mutation probability adjustments are effective enhancements for mutation-based neuroevolution.
• These methods offer a robust alternative to gradient-based approaches for ANN design.
• The study highlights the potential of refined mutation strategies in advancing neuroevolutionary algorithms.