Normalisation

Submodules

Normalisation base class

class fleetrl.utils.normalization.normalization.Normalization[source]

Bases: object

Parent class for Normalization.

make_boundaries(dim)[source]

Create the observation space for gym. Depending on the observation, upper and lower boundaries can change.

Parameters:: dim (tuple[int]) – The length of the array, depends on the number of cars, and other information contained in the obs
Return type:: tuple[float, float] | tuple[ndarray, ndarray]
Returns:: A tuple, containing the low obs and high obs array that will be parsed to gym.Spaces.box

normalize_obs(obs)[source]

Normalizes the values in an observation.

Parameters:: obs (dict) – An observation, containing the state of charge, hours left, and price
Return type:: ndarray
Returns:: The normalized observation, concatenated to a single array

Oracle normalisation

class fleetrl.utils.normalization.oracle_normalization.OracleNormalization(db, building_flag, pv_flag, price_flag, ev_conf, load_calc, aux)[source]

Bases: Normalization

Oracle Normalization assumes the knowledge of the max and min values of the dataset. This is necessary to perform a global min/max normalization. Alternatively, a rolling normalization could be implemented.

static flatten_obs(obs)[source]

Observations must be flattened for openAI gym compatibility. The parsed observation must be a flat array and not a dictionary. The dictionary either includes float or array. The function removes the nesting.

Parameters:: obs – Normalized observation dictionary
Returns:: A flattened array - necessary for the RL algorithms to be in a 1-dim array e.g. [v_1, …, v_N]

make_boundaries(dim)[source]

The boundaries are 0 and 1 because the observations are min/max normalized.

Parameters:: dim (tuple[int]) – Dimension of the observation depending on the flags
Return type:: tuple[float, float] | tuple[ndarray, ndarray]
Returns:: Low and high observation arrays for gym.Spaces.

normalize_obs(input_obs)[source]

Normalization function. Different cases are checked depending on the flags of PV, load, price, and aux. Input observations are a dictionary with clear namings, to make further changes in the code easy and readable.

Parameters:: input_obs (dict) – Input observation: Un-normalized observations as specified in the observer.
Return type:: ndarray
Returns:: Normalized observation.

Unit normalisation

class fleetrl.utils.normalization.unit_normalization.UnitNormalization[source]

Bases: Normalization

This function does not normalize, but parses the un-normalized values.

static flatten_obs(obs)[source]

Flattening nested dict to np.ndarray

Parameters:: obs – Observation dictionary including floats or arrays
Returns:: Flattened observation

make_boundaries(dim)[source]

The bounds can be -inf and inf. This is the least restricting path and allows for cross-compatibility of agents on environments that have the same shape.

Parameters:: dim (tuple[int]) – Dimension depending on the flags.
Return type:: tuple[float, float] | tuple[ndarray, ndarray]
Returns:: Low_obs and high_obs for gym.Spaces.

normalize_obs(obs)[source]: Obs is flattened and returned. :type obs: dict :param obs: Observation dictionary that includes floats or array :rtype: ndarray :return: Flattened np.ndarray