Reinforcement Learning for the Detection of Arbitrary Shaped Spatial Clusters

Investigations of spatial clustering patterns arouse interest in several areas. The most widely used technique for this objective is the Spatial Scan Statistic. However, the Scan statistics alone are ineffective when clusters may have an irregular shape. Other methodologies are used to increase flexibility for possible cluster shapes. This article proposed an approach through Reinforcement Learning with Scan spatial statistics as the reward function. The new methodology is presented in detail and tested in a vast controlled experimental set to ensure its adaptability to the problem of detecting spatial clusters. The results show that the new technique is promising for the problem under study.

I. Introduction

Investigations to understand disease occurrence is an essential need in several everyday applications [1]. The count of these occurrences may arouse interest when they are excessively higher or lower than the previously expected average [2]. Several areas of study require the use of techniques for this purpose; applications in criminology [3], ecology [4], and epidemiology [5], among others, are commonly observed in the literature. In particular, interest in epidemiology and syndromic surveillance studies is associated with occurrence in human populations. Diseases cases can often present higher than expected quantities. In this case, public management must carry out mitigation actions in the most agile and assertive way. Given this, the location of discrepant regions for occurrences is a valuable resource for carrying out corrective measures [6].