Adaptive Normal Two-Armed Bandit and Data Processing Optimization*

We consider the two-armed bandit problem as applied to data processing provided that there are two alternative processing methods with different a priori unknown efficiencies. One should determine more efficient method and ensure its preferable application. Normal two-armed bandit is a generalization which allows to process data in parallel and almost without loss of the control performance, i.e. without increasing of the minimax risk. However, it requires that methods must have close efficiencies. Below we propose the adaptive modification of the algorithm which works properly with methods which efficiencies are not obligatory close.