BanditPolicy

Scoring strategy for a com.eignex.kumulant.bandit.univariate.MultiArmedBandit. Decides which arm to play given snapshots of each arm's sufficient statistic R. The bandit calls evaluate for every arm and picks the argmax; the policy is the entire exploration/exploitation knob.

The policy owns the per-arm cumulator lifecycle through its arm spec:

• createArm returns a freshly-prior-seeded SeriesStat from arm.createStat().

• update folds an observation in, applying arm.encode first so the stat sees the encoded value (e.g. ln(value) for LogNormalArm).

• evaluate reads the resulting snapshot.

Two flavours:

• Sampling-based (ThompsonSampling): score each arm by a draw from its conjugate Posterior given the snapshot. Exploration is implicit in posterior variance: under-explored arms have wider posteriors and draw higher scores more often.

• UCB-based (UCB1, UCB1Normal, UCB1Tuned, UcbV, KlUcb, Moss) ; score is mean + alpha * confidence-bound derived from the snapshot directly. Exploration is explicit in the confidence width.

Per-policy global state (e.g. total samples for UCB) updates through addArm / removeArm when the arm population changes mid-run, and through update's side effects on each observation.

Inheritors