Implementation of tabular methods in reinforcement learning for policy control and evaluation.
We demonstrate different policy evaluation (Monte Carlo, k step Temporal Difference) and control (k step SARSA, TD lambda, Monte Carlo) algorithms for the game of BlackJack.
blackjack_simulator.py is an implementation of BlackJack simulator. It also defines an abstract class for Environment. Any suitable class derived from this abstract class can be substituted in place of BlackJack to run the RL algorithms for a different setting.
evaluate_policy.py contains implementation for different policy evaluation algorithms while control_policy.py contains policy control algorithms. It also contains test_policy method to evaluate any policy by computing average rewards by acting out in the environment according to the given policy.
tabular_rl.py implements a wrapper over the implemented algorithms providing a clean interface for benchmarking the algorithms.
For a more interactive guide and visualization of algorithms, and observing the effect of different hyperparameters refer to BlackJackDemo.ipynb notebook.