Deep Reinforcement Learning for Spatial Data Partitioning

A framework for applying deep reinforcement learning to spatial data partitioning.

Requirements

• JDK 1.8
• Spark 2.3.4
• Hadoop 2.7.7
• GeoSpark 1.3.1 (Apache Sedona)

Code

Execute the following commands on the master node (or locally).

$ python python/main.py

To run main-training, you need to prepare jar files compiled a core of Sedona /sedona/geospark-1.3.1.jar and a scala program for run queries /sedona/geosparkapplication_2.11-0.1.jar. In addition, we have also modified Sedona jar file of open source so that obtained during training partitions /tmp/partitions.csv can be read externally. In the last line of python/config.yaml, the --master address and the path to these jar files must be set appropriately.

Dataset

You can execute program using the two datasets OSM-US, OSM-SA. Or you can use any dataset that consists of two columns of latitude and longitude.