Currently, pgmpy deals with only discrete random variables. In many situations, some variables are best modeled as taking values in some continuous space. Examples include variables such as position, velocity etc.
The first part of the project creates a module to represent nodes having a continuous domain representation. These nodes would be used in hybrid networks comprising both continuous as well as discrete random variables. The two important features in this part would be -
- Representation of User Defined Continuous Random Variables
- Methods to convert continuous distributions into discrete factors.
The second part of the project will deal with Gaussian distributions. Gaussians are a particularly simple subclass of distributions that make very strong assumptions, such as the exponential decay of the distribution away from its mean, and the linearity of interactions between variables. Gaussians are a surprisingly good approximation for many real world distributions.
There will be support for variables comprising the most popular forms of representation in Gaussian distributions -
- Linear Gaussian Distribution
- Joint Gaussian Distribution
- Canonical Forms