6533b85bfe1ef96bd12bad13
RESEARCH PRODUCT
DOBRO : a prediction error correcting robot under drifts
Eric BouilletTommi KärkkäinenAlexandr V. MaslovHoang Thanh LamMykola Pechenizkiysubject
ta113Concept driftComputer scienceMean squared prediction error02 engineering and technologyARIMAconcept drifton-line prediction error correction020204 information systems0202 electrical engineering electronic engineering information engineeringRobot020201 artificial intelligence & image processingAutoregressive integrated moving averageSimulationdescription
We propose DOBRO, a light online learning module, which is equipped with a smart correction policy helping making decision to correct or not the given prediction depending on how likely the correction will lead to a better prediction performance. DOBRO is a standalone module requiring nothing more than a time series of prediction errors and it is flexible to be integrated into any black-box model to improve its performance under drifts. We performed evaluation in a real-world application with bus arrival time prediction problem. The obtained results show that DOBRO improved prediction performance significantly meanwhile it did not hurt the accuracy when drift does not happen.
| year | journal | country | edition | language |
|---|---|---|---|---|
| 2016-04-04 |