DOBRO: a prediction error correcting robot under drifts

摘要: We propose DOBRO, a light online learning module, which is equipped with smart correction policy helping making decision to correct or not the given prediction depending on how likely will lead better performance. DOBRO standalone module requiring nothing more than time series of errors and it flexible be integrated into any black-box model improve its performance under drifts. performed evaluation in real-world application bus arrival problem. The obtained results show that improved significantly meanwhile did hurt accuracy when drift does happen.