Demonstrating real advantage of machine learning-enhanced Monte Carlo for combinatorial optimization.

Area of Science:

• Computational Physics
• Machine Learning
• Optimization

Background:

• Combinatorial optimization is crucial for applications and method development.
• Machine learning-assisted optimization is emerging but often underperforms classical methods.
• Quadratic Unconstrained Binary Optimization (QUBO) problems, like Ising spin glasses, are challenging.

Purpose of the Study:

• To develop and evaluate a novel machine learning-assisted optimization algorithm for QUBO problems.
• To investigate the performance of a Global Annealing Monte Carlo method against classical algorithms.
• To demonstrate the effectiveness of integrating machine learning-proposed global moves with local search.

Main Methods:

• Developed a Global Annealing Monte Carlo algorithm incorporating machine learning-guided global moves.
• Applied the algorithm to find minimum energy configurations in 3D Ising spin glasses.
• Benchmarked performance against Simulated Annealing and Population Annealing.

Main Results:

• The Global Annealing algorithm significantly outperformed Simulated Annealing.
• It demonstrated greater robustness than Population Annealing across varying problem hardness and system sizes.
• The integration of local moves was found to be critical for optimal performance.

Conclusions:

• Machine learning-assisted optimization, specifically Global Annealing, can surpass state-of-the-art classical methods for combinatorial optimization.
• The developed method shows robustness and effectiveness without requiring hyperparameter tuning.
• This work highlights the potential of hybrid machine learning approaches in tackling complex optimization challenges.