Q-Learner

A tabular Q-learning agent dropped into a 2D gridworld with nothing but a reward signal and an ε-greedy disposition. It wanders, it bumps into walls, and — after enough episodes — it stops embarrassing itself.

The interesting part was reward shaping: small tweaks to the signal changed the personality of the policy more than any hyperparameter. Convergence plots and a value heatmap document the journey from random walk to competence.

• PYTHON + NUMPY
• ε-GREEDY POLICY
• REWARD SHAPING
• VALUE HEATMAP