Project case study · Reinforcement learning
Q-Learning on FrozenLake
A focused implementation of tabular Q-learning that makes exploration, temporal-difference updates, policy extraction, and evaluation explicit rather than hiding the learning loop behind a high-level reinforcement-learning library.
Implementation
- Initialize the Q-table and calculate maximum action values.
- Choose actions with greedy selection and epsilon-greedy exploration.
- Decay epsilon over training.
- Compute temporal-difference targets and errors.
- Update Q-values from environment interactions.
- Extract a greedy policy and evaluate episode success rate.
Learning loop
Observe state → choose an action → receive reward and next state → compute the TD target → update the state-action value → repeat across training episodes.
Run and evaluate
The repository documents running the implementation with python scaffold.py. Success rate depends on environment settings, training configuration, and evaluation episodes; this page does not claim a specific result without a reproducible report.