Robust Data Sampling in Machine Learning: A Game-Theoretic Framework for Training and Validation Data Selection

How to sample training/validation data is an important question for machine learning models, especially when the dataset heterogeneous and skewed. In this paper, we propose a sampling method that robustly selects data. We formulate process as two-player game: trainer aims training so minimize test error, while validator adversarially samples validation can increase error. Robust achieved at game equilibrium. To accelerate searching process, adopt reinforcement aided Monte Carlo trees search (MCTS). apply our car-following modeling problem, complicated scenario with random human driving behavior. Real-world data, Next Generation SIMulation (NGSIM), used validate method, experiment results demonstrate robustness thereby model out-of-sample performance.