← Back to project summary

Q-Learning Gridworlds

A closer look at the app, the systems I owned, and the features behind it.

What is in the repo

  • Environment setup for Taxi and for Frozen Lake in both deterministic and slippery variants, with the time limit wrapper rewrapped so episode length is controlled explicitly.
  • An epsilon-greedy training loop and a separate greedy evaluation loop that reports average discounted reward.
  • Debugging aids that print the learned policy as a readable grid.
  • Experiment sweeps with pandas and Plotly for charting learning rate and epsilon.

What is missing

  • The QValues data structure, the Q-learning update rule, the value lookup, and best-action selection are all left as TODO.
  • Listed here for completeness, and because the harness around it is a reasonable piece of work on its own.