Uncertainty-Aware Data Aggregation for Deep Imitation Learning

摘要: Estimating statistical uncertainties allows autonomous agents to communicate their confidence during task execution and is important for applications in safety-critical domains such as driving. In this work, we present the uncertainty-aware imitation learning (UAIL) algorithm improving end-to-end control systems via data aggregation. UAIL applies Monte Carlo Dropout estimate uncertainty output of systems, using states where it uncertain selectively acquire new training data. contrast prior aggregation algorithms that force human experts visit sub-optimal at random, can anticipate its own mistakes switch expert order prevent visiting a series states. Our experimental results from simulated driving tasks demonstrate our proposed estimation method be leveraged reliably predict infractions. analysis shows outperforms existing on benchmark tasks.