Reinforcement Learning for Multi-Step Expert Advice

Complex tasks for heterogeneous data sources, such as finding and linking named entities in text documents or detecting objects in images, often require multiple steps to be solved in a processing pipeline. In most of the cases, there exist numerous, exchangeable software components for a single step, each an "expert" for data with certain characteristics. Which expert to apply to which observed data instance in which step becomes a challenge that is even hard for humans to decide. In this work, we treat the problem as Single-Agent System (SAS) where a centralized agent learns how to best exploit experts. We therefore define locality-sensitive relational measures for experts and data points, so-called "meta-dependencies", to assess expert performances, and use them for decision-making via Online Model-Free-and Batch Reinforcement Learning (RL) approaches, building on techniques from Contextual Bandits (CBs) and Statistical Relational Learning (SRL). The resulting system automatically learns to pick the best pipeline of experts for a given set of data points. We evaluate our approach for Entity Linking on text corpora with heterogeneous characteristics (such as news articles or tweets). Our empirical results improve the estimation of expert accuracies as well as the out-of-the-box performance of the original experts without manual tuning.