Description
There are several existing classes that appear to be reproducing closely related functionalities, thus violating the "Don't Repeat Yourself" principle which is outlined in one of the lectures in one of the lectures. For instance, all the following modules include a form of a linear state transition equation
robustlq.py
lss.py
lqnash.py
lqcontrol.py
army.py
kalman.py
#476 introduces yet another class which falls into this group. The ARMA
class has a simulation
method, the LQ
class has a compute_sequence
method and the LinearStateSpace
has a simulate
method. Besides, both Kalman
and LQ
have stationary_values
methods that use solve_discrete_riccati
. Similarly, LinearStateSpace
and ARMA
have impulse_response
methods.
Therefore, we might want to consider refactoring these classes, potentially into a sub-package. Besides avoiding repetition and making it easier to understand how these classes are connected, another benefit would be that features added to a base class would directly propagate to derived classes.
As @thomassargent30 pointed out, this refactoring is bound to take some time and require a person who is familiar with all of the lectures that use these classes. Considering that some new lectures which use these classes are currently in production, @jstac suggested that we hold off on a major reorganization and instead consider piecemeal improvements. Specifically, we could add a linear_gaussian
module that contains:
-
a jitted function that simulates vector-valued Gaussian processes, and
-
a jitted function that returns the stationary mean and variance.
Does anyone have any additional thoughts or suggestions?