Partially Observable Reinforcement Learning for Sustainable Active Surveillance

Abstract

Active surveillance is the most effective strategy in the applications of infectious disease prevention, road network optimization, crime reconnaissance, etc. However, the incomplete data collected from partially monitored regions by active surveillance disables existing models to maintain a sustainable performance in the future. To address this issue, this article presents a sustainable active surveillance framework (SAS), which consists of a predictor, a classifier, and a planner, by developing a novel partially observable reinforcement learning algorithm. The predictor estimates priorities of candidate regions for monitoring. The classifier assigns candidate regions with similar features into the same groups, so that the data collected from monitored regions can be shared with unmonitored regions within the group. The planner determines where and when to allocate limited resources, considering the outcomes of available resources and model sustainability. An empirical case study on infectious disease prevention showed that the proposed SAS method significantly outperforms the state-of-the-art methods.

This work was funded by the National Natural Science Foundation of China under grants 61373053 and 61572226, and in part by the Jilin Province Key Scientific and Technological Research and Development Project under Grant 20180201067GX and Grant 20180201044GX.