Filter and forecasting

• These are example script-based runs to filter observed data using cd-dynamax, based on a given filter (as specified via --filter_config)

  • Kalman Filtering and forecasting with a continuous-discrete Gaussian State-space model data

  python3 run_filter_then_forecast.py --data_config_file data/true_cdlgssm_data_x1 --model_config_file model/cdlgssm_x1 --filter_config_file filter/kf
  

  • Filtering and forecasting a continuous-discrete Lorenz 63 model with EKF and EnKF filters

  python3 run_filter_then_forecast.py --filter_config_file filter/ekf_StateFirst_EmissionsFirst
  

  python3 run_filter_then_forecast.py --filter_config_file filter/ekf_StateFirst_EmissionsFirst --data_key 0 1 2
  

  python3 run_filter_then_forecast.py --filter_config_file filter/enkf_StateFirst --ftf_key 0 1 2
  

  • An example to run all (pre-defined) CD-NLGSSM filters in one simulation

  python3 run_filter_then_forecast.py --filter_config_file all --data_key 10 --ftf_key 10
  

• An example of a Differentiable Particle Filter for a CD-NLSSM model

python3 run_filter_then_forecast.py --data_config_file data/true_oudrift_poissondata --model_config_file model/cdnlssm_oudrift_poissondata --filter_config_file filter/dpf_bootstrap_stopg

Plotting the filter and forecast results

• To plot results from the above scripts, run the plot_filter_then_forecast.py with the same config files as when using `run_filter_then_forecast.py script

  • For instance, to plot results for the KF run above

  python3 plot_filter_then_forecast.py --data_config_file data/true_cdlgssm_data_x1 --model_config_file model/cdlgssm_x1 --filter_config_file filter/kf
  

  • For instance, to plot results for one of the EKFs above run

  python3 plot_filter_then_forecast.py --filter_config_file filter/ekf_StateFirst_EmissionsFirst
  

• To compare all executed CD-NLGSSM filters and plot their results, execute

python3 compare_filter_then_forecast.py --filter_config_file all --data_key 10 --ftf_key 10

• To plot the example of a Differentiable Particle Filter for a CD-NLSSM model

python3 plot_filter_then_forecast.py --data_config_file data/true_oudrift_poissondata --model_config_file model/cdnlssm_oudrift_poissondata --filter_config_file filter/dpf_bootstrap_stopg

Parameter learning: i.e., fitting a cd-dynamax model to observed data

Using SGD

• Run the parameter learning run_fit_model_to_data.py script, with corresponding fit_sgd config file and default filter

  • For a CD-LGSSM model

  python3 run_fit_model_to_data.py --data_config_file data/true_cdlgssm_data_x1 --model_config_file model/cdlgssm_x1_fit_to_data --filter_config_file filter/kf --fit_config_file fitting/fit_sgd_cdlgssm_x1 
  

  • For a Lorenz 63 CD-NLGSSM model with EKF (defaults in script)

  python3 run_fit_model_to_data.py --fit_config_file fitting/fit_sgd
  

  • For a Lorenz 63 CD-NLGSSM model with selected EnKF filter

  python3 run_fit_model_to_data.py --fit_config_file fitting/fit_sgd --filter_config_file filter/enkf_StateFirst
  

• Parameter learning results can be plotted using the plot_fitted_model.py script, with same config files as above, for the SGD-based learning:

  • For the CD-LGSSM model

  python3 plot_fitted_model.py --data_config_file data/true_cdlgssm_data_x1 --model_config_file model/cdlgssm_x1_fit_to_data --filter_config_file filter/kf --fit_config_file fitting/fit_sgd_cdlgssm_x1 
  

  • For the Lorenz63 CD-NLGSSM model with EKF

  python3 plot_fitted_model.py --fit_config_file fitting/fit_sgd
  

Using MCMC - Nuts

• Run the parameter learning run_fit_model_to_data.py script, with corresponding fit_nuts config file

python3 run_fit_model_to_data.py --fit_config_file fitting/fit_nuts

• Parameter learning results can be plotted using the plot_fitted_model.py script, with same config files as above, for the NUTS-based learning

python3 plot_fitted_model.py --fit_config_file fitting/fit_nuts

Scipy

• Run the parameter learning run_fit_model_to_data.py script, with corresponding fit_scipy config file

python3 run_fit_model_to_data.py --fit_config_file fitting/fit_scipy

• Parameter learning results can be plotted using the plot_fitted_model.py script, with same config files as above

python3 plot_fitted_model.py --fit_config_file fitting/fit_scipy