Policy-gradient bandit

Made to be adapted.

Six actions have known deterministic rewards. The agent samples an action from its current policy instead of always choosing the best one. The seed makes the sequence reproducible.

A six-outcome mathematical model with exact updates. No language model is being trained.