Limitations of parallel global optimization for large-scale human movement problems.

Area of Science:

• Biomechanics
• Computational modeling
• Optimization algorithms

Background:

• Global optimization algorithms are increasingly used in biomechanics.
• Previous applications were limited to small- to medium-scale problems (<100 variables).
• Parallel computing advances enable larger-scale applications.

Purpose of the Study:

• Evaluate the effectiveness of parallel particle swarm optimization for large-scale human movement problems.
• Compare its performance against gradient-based methods.
• Assess the impact of penalty methods on optimization outcomes.

Main Methods:

• Utilized a dynamic, 27 degree-of-freedom, full-body gait model.
• Applied parallel particle swarm and gradient-based nonlinear least squares algorithms.
• Addressed constraints using a penalty method.
• Two large-scale problems (660 variables) focused on minimizing knee adduction torque.

Main Results:

• Gradient-based nonlinear least squares significantly outperformed particle swarm optimization.
• A single gradient-based run found better solutions than 10 particle swarm runs.
• Penalty terms created a narrow design space channel difficult for global optimizers to enter.

Conclusions:

• Researchers should be cautious when applying parallel global optimizers to large-scale human movement problems.
• Gradient information is crucial for navigating complex design spaces with penalty constraints.
• Particle swarm optimization may not be suitable for these specific large-scale biomechanical challenges.