config_dmbrl module

Low level configuration for modeling and optimisation.

  • ctrl_cfg: Configuration parameters for the control algorithm.

    -dO: dimensionality of observations -dU: dimensionality of control inputs - per: How often the action sequence will be optimized, i.e, for per=1 it is reoptimized at every call to MPC.act(). - constrains: [[np.array([min v, min q]), np.array([max v, max q])], [min q/v, max q/v], [min q/sqrt(v), max q/sqrt(v)]] - prop_cfg: Configuration parameters for modeling and uncertainty propagation.

    • model_pretrained: True if model used for MPC has been trained on previous data, False otherwise.

    • model_init_cfg: Configuration parameters for model initialisation.

      • ensemble_size: Number of models within the ensemble.
      • load_model: True for a pretrained model to be loaded upon initialisation.
      • model_dir: Directory in which the model files (.mat, .nns) are located.
      • model_name: Name of the model files (model_dir/model_name.mat or model_dir/model_name.nns)
    • model_train_cfg: Configuration parameters for model training optimisation
      • batch_size: Batch size.
      • epochs: Number of training epochs.
      • hide_progress: If ‘True’, additional information regarding model training is printed.

    • npart: Number of particles used for uncertainty propagation.

    • model_in: Number of inputs to the model.

    • model_out: Number of outputs to the model.

    • n_layers: Number of hidden layers.

    • n_neurons: Number of neurons per hidden layer.

    • learning_rate: Learning rate.

    • wd_in: Weight decay for the input layer neurons.

    • wd_hid: Weight decay for the hidden layer neurons.

    • wd_out: Weight decay for the output layer neurons.

    • opt_cfg: Configuration parameters for optimisation.

      • mode: Uncertainty propagation method.

      • plan_hor: Planning horizon for the model predictive control algorithm.

      • cfg

        • popsize: Number of cost evaluations per iteration.
        • max_iters: Maximum number of optimisation iterations.
        • num_elites: Number of elites.
        • alpha: Alpha parametero of the CEM optimisation algorithm.
        • eps: Epsilon parameter of the CEM optimisation algorithm.

      • prop_cfg

        • mode: Uncertainty propagation method, ie “TSinf”

    • change_target: True if multiple setpoints used, i.e. 980 and 1010

    • n_parts_targets: Number of parts to be built for each target

    • targets: Different temperature setpoints to be used (must be of same length as n_parts_targets)

    • force: Configuration parameters to periodically overwrite (“force”) predefined build parameters

      • on: Force functionality enabled if True
      • start_part: First part where functionality is enabled (disregarding the first few ignored parts)
      • n_parts: Number of parts for which the functionality is enabled
      • n_repeats: Number of consecutive layers for which inputs are forced. For [1,2], n_parts will be forced only once (periodically), while a further n_parts will be forced two times consecutively (periodically)
      • init_buffer: Initial number of layers for which parameters are not forced
      • upper_init: Upper bound is initialised to this.
      • upper_delta: Upper bound increases by this. For instance, for upper_init=105 and upper_delta=5, the upper bound sequence will be 105, 110, 115…
      • lower_init: Lower bound is initialised to this.
      • lower_delta: Lower bound is increased by this. For instance, for lower_init=65 and lower_delta=-5, the lower bound sequence will be 60, 55, 50…
      • fixed_speed: For the forced parameters, power will be adjusted but mark speed will be kept fixed to this value.
config_dmbrl.ac_cost_fn(acs)

Cost function (action-dependent) used in the optimisation problem.

Should process both np.arrays and tf.Tensor inputs.

Parameters:acs (np.array or tf.Tensor) – Array of actions with shape (n, dU).
Returns:float
config_dmbrl.bnn_constructor(model_init_cfg)

Constructs the Bayesian Neural Network model.

Moodel_init_cfg is a dotmap object containing:
  • model_in (int): Number of inputs to the model.
  • model_out (int): Number of outputs to the model.
  • n_layers (int): Number of hidden layers.
  • n_neurons (int): Number of neurons per hidden layer.
  • learning_rate (float): Learning rate.
  • wd_in (float): Weight decay for the input layer neurons.
  • wd_hid (float): Weight decay for the hidden layer neurons.
  • wd_out (float): Weight decay for the output layer neurons.
Returns:BNN class object
config_dmbrl.create_dmbrl_config()

Returns the low-level modeling and optimisation configuration parameters.

config_dmbrl.obs_cost_fn(obs, target)

Cost function (state-dependent) used in the optimisation problem.

Should process both np.arrays and tf.Tensor inputs.

Parameters:
  • obs (np.array or tf.Tensor) – Array of observations with shape (n, dO).
  • target (float) –
Returns:

float
config_dmbrl.obs_postproc(obs, pred)

Modifies observations and model p redictions before being passed to cost function.

Parameters:
  • obs (np.array or tf.Tensor) – Array of observations with shape (n, dO).
  • pred (np.array or tf.Tensor) – Array of predictions (model outputs) with shape (n, model_output).
Returns:

np.array or tf.Tensor with shape (n, dO)
config_dmbrl.obs_preproc(obs)

Modifies observations (in a 2D matrix) before they are passed into the model.

Parameters:obs (np.array or tf.Tensor) – Array of observations with shape (n, dO).
Returns:np.array or tf.Tensor with shape (n, model_in)
config_dmbrl.targ_proc(obs, next_obs)

Takes current observations and next observations and returns the array of targets (so that the model learns the mapping obs -> targ_proc(obs, next_obs))

Parameters:
  • obs (np.array or tf.Tensor) – Array of observations at time t with shape (n, dO).
  • next_obs (np.array or tf.Tensor) – Array of observations at time t+1 with shape (n, dO).
Returns:

np.array or tf.Tensor with shape (n, model_out)