Encouraging behavioral diversity in evolutionary robotics: an empirical study.

Area of Science:

• Robotics
• Artificial Intelligence
• Evolutionary Computation

Background:

• Evolutionary robotics (ER) designs robots and controllers automatically.
• ER often faces premature convergence due to task-specific fitness functions and diverse phenotypes.
• Encouraging behavioral diversity, rather than genotypic diversity, is a recent approach to address this.

Purpose of the Study:

• To review and benchmark published approaches for fostering behavioral diversity in ER.
• To investigate the effectiveness of different behavioral diversity methods across various tasks and genotypes.
• To determine the optimal strategies for enhancing the evolutionary process in ER.

Main Methods:

• A common framework was established to benchmark different behavioral diversity approaches.
• Multiple tasks and genotypes were used to compare the effectiveness of each method.
• Multi-objective methods and Hamming-based behavioral distance were specifically evaluated.

Main Results:

• Fostering behavioral diversity significantly improved the evolutionary process in all tested experiments.
• The effectiveness of behavioral diversity was independent of the specific genotype or task.
• Multi-objective methods demonstrated the highest efficiency among the benchmarked approaches.

Conclusions:

• Behavioral diversity is a robust strategy for improving evolutionary robotics, irrespective of task or genotype.
• Multi-objective optimization and generic behavioral distance metrics are highly effective.
• This research provides a benchmark for future studies on diversity in evolutionary robotics.