search

schuderer / bprl

Business Process Control using Reinforcement Learning. Work in progress.
MIT License
3 stars 0 forks source link

Approximator class for easier swapping out of agents #7

Closed schuderer closed 5 years ago

schuderer commented 5 years ago

This is done. However, the resulting code is ca. 25% slower than the original code (even after some optimization). Therefore, I am hesitant to merge it to master just yet.

It's a bit of a pity to not use this, though, as it quite nicely separates the agent itself from the q-function (with or without approximation) as well as from the exploration/update policy (which can be passed as python functions, which makes the whole thing quite flexible).

The refactored code is in the ref_approximator branch

schuderer commented 5 years ago

The speed issues have been solved, and according to some comparisons the speed is the same as before the refactoring.