q-learning-Taxi-v3

Tabular Q-learning agent trained from a zero Q-table for the Hugging Face Deep RL Course, Unit 2. Prepared for Vlad Kolomiiets (Vladprime).

Training: 100,000 episodes, 1,883,993 environment steps, seed 20261101. No pretrained policy was used.

Independent evaluation: 1,000 episodes with seeds 1000000–1000999. Mean reward 7.8610, standard deviation 2.4931, mean minus standard deviation 5.3679. Full episode rewards are in results.json.

The checkpoint was reloaded and the complete evaluation reproduced exactly. This model is one practical assignment, not a certificate for the whole course.

Course assignment · Certification requirements

Reproduce

pip install -r requirements.txt
python train.py

For inference, use np.load("qtable.npy", allow_pickle=False) and choose int(np.argmax(qtable[state])). The course-format q-learning.pkl is also included. Only load pickle files from trusted sources.