Reinforcement Learning Projects

Solve two classic grid worlds with tabular Q-learning. Inspect FrozenLake's spaces and transition table, measure a random-policy baseline, implement Q-learning with a decaying epsilon-greedy schedule, and watch the success rate climb. Visualize the learned policy as arrows over a value-function heatmap, evaluate the greedy policy on the slippery lake, then reuse the exact same agent on Taxi-v4. No downloads, runs in about a minute.

Balance a pole on a moving cart with a Deep Q-Network trained from scratch in PyTorch. Explore CartPole's continuous state and a random baseline, build an experience-replay buffer, a small MLP Q-network with a periodically synced target network, an epsilon schedule and a Huber TD loss, then train under a hard step cap. Finish by plotting the learning curve, evaluating the greedy policy on seeded episodes and tracing the pole angle through a full rollout.