Lab .

Inspect the reward curve, then replay the policy learned by the model.

Choose an environment and compare its observations, actions, and training feedback.

Compare random, greedy, epsilon-greedy and UCB under fixed and changing reward probabilities.

Compare known-model value iteration and interaction-based Q-learning on the same map.

Train DQN and Double DQN with the same budget and separate evaluation seeds.

Enumerate gradient variance, then train policies with and without a learned state baseline.

Train continuous-torque A2C and save both the policy and normalization statistics.

Compile the update function, then inspect training and policy replays.

Train in a small shooting scenario and replay the current policy.

Use frame stacking and a CNN to learn actions from game images.

Train a visual PPO policy inside the Minecraft simulator.

Choose a scene and inspect the behavior learned by PPO.

Start with a small maze and observe exploration under sparse rewards.

Train a shared policy in cooperative and competitive settings.

Start with tic-tac-toe and small poker games to observe interacting policies.

Train only on fixed data and compare naive Q maximization, conservative penalties and behavior constraints.

Train a tiny autoregressive GRU on CPU, then update its parameters using paired preferences.

Check a strict answer verifier and inspect token-level clipping, KL and padding gradients.

Train a tiny autoregressive policy in a restricted tool language, then render calls as JSON.

Continue decisions after tool responses and inspect credit from terminal rewards.

Select a trusted patch, execute Python tests and check separate test inputs.

Use SQL and a calculator over synthetic company data in a multi-turn workflow.

Learn candidate query/evidence selection over a fixed teaching corpus; score answers and citations separately.

Learn point-mass dynamics, measure open-loop drift and plan in the actual environment.

Train a policy in a robot manipulation task and replay its actions.

Navigate an interactive room by choosing actions from visual observations.