Learning through actions
Explore how rewards and experience shape decisions.
What you’ll explore
- Reinforcement learning
- Exploration
- Self-play
Your lessons
0 / 14 completeExplore before you decide
Choose between uncertain rewards and discover why trying a new option can improve later decisions.
Learn from delayed rewards
Teach an agent a route, inspect a Q-learning update, and uncover a reward rule that rewards the wrong behavior.
Solve the Bellman equations
Compute values in a small looping decision process, separate policy evaluation from improvement, and compare exact expectations with sampled returns.
What can the agent observe?
Hide a cart’s velocity, compare aliased observations and recover motion from a short exact history.
Copy a demonstrated action
Train a puzzle policy from exact expert targets, inspect legal masks and compare proposals after a non-expert move.
Learn from verified self-play
Generate game experience, check it against exact rules, and use verified outcomes to change future decisions.
Know which policy produced the data
Compare a logged behavior policy with a target policy using trajectory weights, coverage checks and bounded replay.
Stabilize neural Q-learning
Train a small neural action-value model and compare replay sampling with delayed target networks.
Increase the probability of useful actions
Learn a stochastic policy from sampled returns and inspect the variance of actual updates.
Use a critic to guide learning
Train separate policy and value heads, trace their coupled updates, and audit a biased bootstrap.
Limit a policy update
Compare raw and clipped probability-ratio objectives on the same frozen batch, and test the limits of clipping.
Trade off reward and exploration with entropy
Train a bounded continuous actor and twin critics, and compare task reward with action-distribution entropy.
Learn without new interactions
Fit policies from a fixed log, trace unsupported action values, and test a conservative objective against behavior cloning.
Evaluate learning fairly
Compare complete seed cohorts, data and fitting budgets, frozen checkpoints and task success under a misleading reward bonus.