Reinforcement Learning in Central Banking
Economics | Machine Learning
- Project
- Reinforcement Learning in Central Banking
- Other title
- Learning an Optimal Interest Policy from Experience
Starting from a simple SISO equation with an interest rate pass-through:
![\[\pi_{t} = A\pi_{t-1} + Bi\]](http://jbgrabowski.com/wp-content/ql-cache/quicklatex.com-b02e839ed9d0b899561d59f5fd63130b_l3.png "Rendered by QuickLaTeX.com")
I recast the problem of dynamic control for the central bank into the modeling framework of reinforcement learning. This simple model of the economy was then built into an OpenAI Gym wrapper to allow an reinforcement learning agent to learn the optimal policy rule:
![\[i^\star = \frac{-AB\pi_t}{B^2 + \alpha}\]](http://jbgrabowski.com/wp-content/ql-cache/quicklatex.com-dc07782a2bf0886af7fe03c209adf4b0_l3.png "Rendered by QuickLaTeX.com")
by trial and error. Rewards were computed using a typical quadratic central bank loss function with one policy target and a cost of change parameter. As a first attempt at solving the environment, I used a basic Q-Learning algorithm. The user can define values for A, B, and cost of changing the instrument 
, and can also choose between a discrete or continuous action space.
You can view the whole project here.