Environment

Generally, there are three possible ways to retrieve data from a Cyber-Physical System:

• a pre-recorded data set,
• a data stream from a simulation,
• a data stream from a real CPS.

Environments in Flowcean are a way to describe the possible data sources for the learning and evaluation procedure. They are implemented via classes. A high-level class diagram of the environment classes is provided below.

• Offline Environment: An offline environment has the ability to load a data set. Data is pre-recorded (offline) and saved in a file or in a database. This data is all at once provided to the learning pipeline. This data is fixed and cannot be (meaningfully) changed. The offline environment provides the interface to a data set but is not the data set itself.
• Incremental Environment: An incremental environment has the ability to provide a stream of single data samples or batches of data. This could be provided by iterating over a data set, by an interface to a simulation, or by a live interface to the real Cyber-Physical System. This is often referred to as passive online learning.
• Active Environment: An active environment has the ability to be influenced by the learning algorithm. This is done by receiving an Action. After an action is received the environment will evolve based on the action and the previous state. After a set time interval or a discrete step inside a simulation, the environment can be observed again. Examples are a controlled simulation or a controlled live experiment.

---
title: A high-level class diagram of the environment classes
---
classDiagram
  Environment <|-- OfflineEnvironment
  Environment <|-- IncrementalEnvironment
  Environment <|-- ActiveEnvironment

  class Environment{
    Provides data to a learner.
    + load() -> Self
    + with_transform(transform) -> TransformedEnvironment[Self]
  }
  class IncrementalEnvironment{
    Loads data in an iterative way.
    + collect(n: int) -> DataFrame
  }
  class OfflineEnvironment{
    Loads data only once and in an non-interactive way.
    + get_data() -> DataFrame
  }
  class ActiveEnvironment{
    Loads data in an interactive way
    allowing to act on the environment.
    + act(Action)
    + step()
    + observe() -> Observation
  }

It is possible to apply Transforms to an environment. This is done by applying a transformation (e.g. resampling or normalization) to the DataFrame that the environment provides. As can be seen in the class diagram, the parent class Environment has a method with_transform() which allows to specify the transforms that are applied to an environment.

Depending on the environment class, different Learning Strategies can be applied. An active learning strategy, for example, can only be applied to an ActiveEnvironment.

For more information on the available classes and how environments are implemented in Flowcean, check out the API.