Fast A/B/n Testing: Exact Multi-Policy Comparison via Tree-Coupled Feedback Sharing

Tree-Coupled A/B Testing (TCAB) reduces reward queries in multi-policy comparisons by sharing feedback across coupled policy histories via a predictable tree structure. This method maintains exact standalone finite-horizon trajectory laws for arbitrary history-dependent contextual-bandit policies. It achieves a query complexity of $T + \sum_{t,e} D_{e,t}$ instead of the standard $JT$ outcomes.