An agent wanders for two hundred steps and finally earns a reward. Which of those steps deserve the credit? Reinforcement learning has three classic answers. Here they are, drawn live on a world you can edit.
Tap any empty space to drop a reward for the agents drifting behind this page.
Each loads a map and settings chosen to make one idea obvious. Watch the ribbon under the grid.
Trains fresh Monte Carlo, Q-learning and SARSA(λ) agents on the current map with your settings, then plots discounted return against the best any policy can do.
| Learner | Episodes to 80% of best | Return, last 20 | Share of best |
|---|---|---|---|
| Run the race to fill this table. | |||