Automatic generation of controllers for embodied legged organisms: a Pareto evolutionary multi-objective approach

Area of Science:

• Artificial Intelligence
• Computational Biology
• Evolutionary Computation

Background:

• Evolving controllers for virtual embodied organisms presents computational challenges.
• Balancing solution quality with computational cost is crucial in evolutionary algorithms.

Purpose of the Study:

• To investigate a self-adaptive Pareto evolutionary multi-objective optimization (EMO) approach for controller evolution.
• To demonstrate the trade-off between solution quality and computational cost.
• To compare the proposed algorithm against other EMO and single-objective evolutionary algorithms (EA).

Main Methods:

• Utilized a self-adaptive Pareto EMO algorithm to evolve controllers for virtual embodied organisms.
• Empirically compared the computational cost and solution quality against weighted sum EMO, single-objective EA, and hand-tuned Pareto EMO algorithms.
• Assessed the algorithm's ability to maintain genetic diversity and reduce redundancy.

Main Results:

• The proposed self-adaptive Pareto EMO algorithm incurred significantly less computational cost than the compared algorithms.
• The algorithm successfully produced controllers with diverse locomotion capabilities in a single run.
• Self-adaptation proved highly beneficial in reducing redundancy and maintaining genetic diversity.

Conclusions:

• The self-adaptive Pareto EMO approach offers an efficient method for evolving virtual organism controllers.
• This method reduces evolutionary computational cost while enabling simultaneous exploration of diverse solutions.
• The inherent multi-objectivity naturally maintains genetic diversity and reduces redundancy.