jax_sgmc.scheduler.adaptive_step_size

jax_sgmc.scheduler.adaptive_step_size(burn_in=0, initial_step_size=0.05, stabilization_constant=100, decay_constant=0.75, speed_constant=0.05, target_acceptance_rate=0.02)[source]

Dual averaging scheme to tune step size for schemes with MH-step.

The adaptive step size uses the dual averaging scheme to optimize the acceptance rate, as proposed by [1].

[1] https://arxiv.org/abs/1111.4246

Parameters:

  • burn_in – Initial iterations, in which the step size should be tuned

  • initial_step_size – Initial value of the step size

  • speed_constant – Bigger constant stabilizes adaption against initial iterations

  • decay_constant – Controls decay of learning rate of the step size

  • speed_constant – Weights acceptance ratio statistics

  • target_acceptance_rate – Desired acceptance rate

Returns:

Returns a specific step size scheduler.