Balanced Sampling of a Tagged Training Pool: Formulation and Solvers
Introduction A training pool for a driving planner consists of logged frames annotated with scenario tags, such as lane changes, ramps, cut-ins and stops at red lights. Scenarios that are rare in the logs may nevertheless be important for the model. The pool is therefore resampled before training: frames from under-represented scenarios are repeated, while those from over-represented scenarios may be removed where permitted. Because each frame can carry several tags, repeat factors must be chosen jointly. A frame tagged with both ramp and is_decel, for example, contributes to both scenario counts, so changing one target affects the other. Assigning one weight to each tag combination allows the scenario targets to be met while minimising changes to the original pool. In survey statistics, this procedure is known as raking, or calibration of sampling weights to known marginal totals [1], [2]. Related work addresses calibration of design weights [3] and generalized raking for regression in two-phase samples [4]. ...