Poster
in
Workshop: Bayesian Decision-making and Uncertainty: from probabilistic and spatiotemporal modeling to sequential experiment design
Adaptive Transductive Inference via Sequential Experimental Design with Contextual Retention
Tareq Si Salem
Keywords: [ Sequential experimental design ] [ Active Learning ] [ Multi-armed Bandits ] [ Online Learning ] [ Decision-making under uncertainty ]
This paper presents a three-stage framework for active learning, encompassing data collection, model retraining, and deployment phases. The framework's primary objective is to optimize data acquisition, data freshness, and model selection methodologies. To achieve this, we propose an online policy with performance guarantees, ensuring optimal performance in dynamic environments. Our approach integrates principles of sequential optimal experimental design and online learning. Empirical evaluations validate the efficacy of our proposed method in comparison to existing baselines.