RBC reinforcement experiment

This is an experiment to examine the use of reinforcement learning to learn an optimal policy in a real-business cycle model. It is part of an effort to apply reinforcement learning to the broader class of macroeconomic DSGE-type models.

Usage

This is still quite rough.

1. Setup the hyperparameter grid-space in experiments_params.yaml.
2. Run Snakemake INIT to build a hyperparameter grid
3. Run Snakemake TRAIN to train each model in the hyperparameter grid

I'm still tweaking this, but in a short-episode-length setting, a PPO agent can learn a policy that outperforms the optimum constant-value policy.

Requirements:

This model was developed in the conda environment descsribed in conda.yaml

Most crucial dependencies are:

• tensorforce (latest)
• tensorflow 1.4
• snakemake (latest)
• gym (latest)