Skip to content

1) Scripts usage

Jump to bottom
Enric Ribera Borrell edited this page Jun 8, 2022 · 2 revisions

Monte Carlo Sampling

sample not controlled trajectories

python src/sampling/sample_not_controlled.py \
  --d 3 \
  --alpha-i 1. \
  --beta 1. \
  --dt 0.01 \
  --K 1000

numerical solution of the HJB equation

solve associated BVP by using finite differences

python src/hjb/compute_hjb_solution.py \
  --d 2 \
  --alpha-i 5. \
  --beta 1. \
  --h-hjb 0.005

Metadynamics algorithm

sample metadynamics trajectories

python src/sampling/sample_metadynamics.py \
  --d 4 \
  --alpha-i 5. \
  --beta 1. \
  --meta-type cum \
  --weights-type const \
  --omega-0 1. \
  --dt-meta 0.01 \
  --K-meta 1 \
  --sigma-i-meta 0.5 \
  --delta-meta 1 \
  --seed 1

Importance Sampling

"Optimal" controlled Sampling

sample optimal controlled trajectories (control from the hjb-solution)

python src/sampling/sample_optimal_controlled.py \
  --d 2 \
  --alpha-i 5. \
  --beta 1. \
  --h-hjb 0.005 \
  --dt 0.01 \
  --K 1000

controlled Sampling with the bias potential from metadynamics. Control represented with uniform distributed gaussian gradients.

sample controlled trajectories (control from the metadynamics algorithm)

python src/sampling/sample_meta_controlled.py \
  --d 1 \
  --alpha-i 5. \
  --beta 1. \
  --meta-type cum \
  --weights-type const \
  --omega-0 1. \
  --dt-meta 0.01 \
  --K-meta 1 \
  --sigma-i-meta 0.5 \
  --sigma-meta 1 \
  --seed 1 \
  --distributed uniform \
  --theta meta \
  --dt 0.01 \
  --K 1000

controlled Sampling with the bias potential from metadynamics. Control represented by feed-forward NN.

sample controlled trajectories (control from the metadynamics algorithm)

python src/sampling/sample_meta_controlled_nn.py \
  --d 1 \
  --alpha-i 5. \
  --beta 1. \
  --meta-type cum \
  --weights-type const \
  --omega-0-meta 1. \
  --dt-meta 0.01 \
  --delta-meta 1 \
  --K-meta 1 \
  --dt 0.01 \
  --K 1000 \
  --d-layers 30 30 \
  --activation tanh \
  --distributed meta \
  --theta meta

Reduced SOC Problem. IPA inexact gradient estimator.

Control represented by Gaussian gradients. SGD with constant learning rate

  • with null initial bias potential
python src/soc/sgd_grad_loss_gaussian_ansatz.py \
  --d 1 \
  --alpha-i 1. \
  --beta 0.5 \
  --distributed uniform \
  --sigma-i 0.5 \
  --m-i 50 \
  --theta null \
  --dt 0.01 \
  --K 1000 \
  --lr 0.01 \
  --n-iterations-lim 10
  • with initial bias potential fitted from the metadynamics algorithm
python src/soc/sgd_grad_loss_gaussian_ansatz.py \
  --d 1 \
  --alpha-i 1. \
  --beta 0.5 \
  --distributed uniform \
  --sigma-i 0.5 \
  --m-i 50 \
  --theta meta \
  --meta-type cum \
  --weights-type const \
  --omega-0-meta 1. \
  --dt-meta 0.01 \
  --sigma-meta 1. \
  --K-meta 1 \
  --dt 0.01 \
  --K 1000 \
  --lr 0.01 \
  --n-iterations-lim 10

Control represented by Gaussian Ansatz Pytorch model. SGD with constant learning rate or Adam

  • nn initialized with random coefficients
python src/soc/sgb_eff_loss_gaussian_ansatz.py \
  --d 1 \
  --alpha-i 1. \
  --beta 1. \
  --theta random \
  --dt 0.01 \
  --K 1000 \
  --optimizer sgd \
  --lr 0.01 \
  --n-iterations-lim 10 \
  --seed 1
  • nn initialized with coefficients trained such that the control is zero in the whole domain
python src/soc/sgb_eff_loss_gaussian_ansatz.py \
  --d 1 \
  --alpha-i 1. \
  --beta 1. \
  --theta null \
  --d-layers 30 30 \
  --activation tanh \
  --dt 0.01 \
  --K 1000 \
  --optimizer sgd \
  --lr 0.01 \
  --n-iterations-lim 10 \
  --seed 1
  • nn initialized with coefficients trained such that the control is the meta controlled
python src/soc/sgb_eff_loss_gaussian_ansatz.py \
  --d 1 \
  --alpha-i 1. \
  --beta 1. \
  --theta meta \
  --d-layers 30 30 \
  --activation tanh \
  --dt 0.01 \
  --K 1000 \
  --optimizer sgd \
  --lr 0.01 \
  --n-iterations-lim 10 \
  --seed 1

Control represented by feed-forward NN. SGD with constant learning rate

  • nn initialized with random coefficients
python src/soc/sgb_eff_loss_feed_forward_nn.py \
  --d 1 \
  --alpha-i 1. \
  --beta 1. \
  --theta random \
  --d-layers 30 30 \
  --activation tanh \
  --dt 0.01 \
  --K 1000 \
  --optimizer sgd \
  --lr 0.01 \
  --n-iterations-lim 10 \
  --seed 1
  • nn initialized with coefficients trained such that the control is zero in the whole domain
python src/soc/sgb_eff_loss_feed_forward_nn.py \
  --d 1 \
  --alpha-i 1. \
  --beta 1. \
  --theta null \
  --d-layers 30 30 \
  --activation tanh \
  --dt 0.01 \
  --K 1000 \
  --optimizer sgd \
  --lr 0.01 \
  --n-iterations-lim 10 \
  --seed 1
  • nn initialized with coefficients trained such that the control is the meta controlled
python src/soc/sgb_eff_loss_feed_forward_nn.py \
  --d 1 \
  --alpha-i 1. \
  --beta 1. \
  --theta meta \
  --d-layers 30 30 \
  --activation tanh \
  --dt 0.01 \
  --K 1000 \
  --optimizer sgd \
  --lr 0.01 \
  --n-iterations-lim 10 \
  --seed 1

Clone this wiki locally