Port-Hamiltonian Optimization

How it works

Benefits

Comparisons

Highlighted claims

• The continuous-time template includes canonical coupling, learned skew dynamics, learned positive semidefinite dissipation, and bounded port input. — When Descent Is Too Stable: Event-Triggered Hamiltonian Learning to Optimize
• The local Hamiltonian combines the base objective, shaped potential, and kinetic energy. — When Descent Is Too Stable: Event-Triggered Hamiltonian Learning to Optimize
• The implementation uses a semi-implicit first-order update for momentum and then position. — When Descent Is Too Stable: Event-Triggered Hamiltonian Learning to Optimize
• The port-Hamiltonian structure separates landscape shaping, energy transport, dissipation, and active energy injection. — When Descent Is Too Stable: Event-Triggered Hamiltonian Learning to Optimize
• Classical gradient-based optimizers are reviewed through a port-Hamiltonian lens. — When Descent Is Too Stable: Event-Triggered Hamiltonian Learning to Optimize
• The energy-balance form provides diagnostics for noise, port work, and discretization defects. — When Descent Is Too Stable: Event-Triggered Hamiltonian Learning to Optimize