Class Ucb1Selector

java.lang.Object
com.quantfinlib.execution.Ucb1Selector
public final class Ucb1Selector extends Object
UCB1 multi-armed bandit — principled selection among venues, LPs or algo variants when the scorecards are still THIN. The exploration problem is real: always using the best-so-far venue means never learning whether another improved; rotating uniformly wastes flow on known-bad ones. UCB1 (Auer et al. 2002) picks the arm maximizing 
  mean reward + √(2·ln N / nᵢ)
— the optimism bonus shrinks as an arm is tried, so exploration decays exactly as fast as the evidence accumulates, with logarithmic regret guaranteed. Rewards must be in [0, 1] (the theory's scale — map fill quality, negated cost bps, or markout onto it; the gate enforces it because a mis-scaled reward silently breaks the exploration balance).

Where this sits vs the scorecards: VenueScorecard and fx.LpScorecard are the RIGHT tool once hundreds of fills per venue exist — they model fill rate, latency and markout separately. UCB1 is for the cold start and for A/B-ing ALGO variants (is the new schedule actually better?), where a single scalar reward and a regret guarantee beat a half-warmed-up model. Deterministic (ties break to the lowest index), O(arms) per selection, allocation-free after construction. Research/warm lane.

  • Constructor Summary

    Constructors
    Constructor
    Description
    Ucb1Selector(int arms)
     

  • Method Summary

    Modifier and Type
    Method
    Description
    double
    meanReward(int arm)
    The arm's observed mean reward (NaN before its first pull).
    long
    pulls(int arm)
    Times an arm has been used.
    void
    record(int arm, double reward)
    Records the observed reward for an arm.
    int
    select()
    The arm to use next: each arm once first (in index order), then highest upper confidence bound, ties to the lowest index.
    long
    totalPulls()
     

    Methods inherited from class java.lang.Object

    clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

  • Constructor Details

    • Ucb1Selector

      public Ucb1Selector(int arms)
      Parameters:
      arms - number of venues/variants, ≥ 2

  • Method Details

    • select

      public int select()
      The arm to use next: each arm once first (in index order), then highest upper confidence bound, ties to the lowest index.

    • record

      public void record(int arm, double reward)
      Records the observed reward for an arm.
      Parameters:
      arm - the arm that was used
      reward - in [0, 1] — the UCB1 theory's scale; rescale upstream, never here

    • pulls

      public long pulls(int arm)
      Times an arm has been used.

    • meanReward

      public double meanReward(int arm)
      The arm's observed mean reward (NaN before its first pull).

    • totalPulls

      public long totalPulls()