Sequential Decision Lab / equation → policy → evidence
Explore how decisions are learned.
From certified Bellman planning to recurrent navigation and quantile critics. Compare measured policies, inspect every training seed, and see what changes when observations, coverage or the learning objective changes.
Inspect the selected policy
Compare methods on this task
Each fitted policy is scored on 100 declared test resets. Training SD is across three fitted policies. Episode intervals condition on one fitted policy. Lecture studies use their declared independent instance counts.
| Method | Fits / instances | Mean score | Training SD | Worst / best fit |
|---|
Bellman sweeps in the stochastic maze
Change the discount and iteration count. Values start at zero; the greedy arrows and contraction certificate are calculated in your browser. This uses the known transition kernel, independently of the trained maze policies.
Energy, packet loss and fairness
Every point is a solved discounted occupancy LP. Channel interpretation, reward priorities and service constraints change the task. These are exact model-aware planning references.
Vertical axis: normalized weighted packet loss.