RLTrajPart

RLTrajPart is a reinforcement learning driven elastic partitioning framework for distributed trajectory stream analytics.

System requirements

• Java 8
• Apache Flink 1.17
• Kafka 3.4.0
• Python.version = 3.8.19
• PyTorch.version = 2.3.0

The system contains two main components: the trajectory stream analytics on the Flink side, and the reinforcement learning agent implemented with PyTorch.

For the Flink side, compile the Java code with mvn package. For the RL agent, install the required Python packages in RL/requirements.txt.

Run

The parameters are configured in TrajectoryFlink\src\main\resources\conf\trajectoryflink-conf.yml and RL/config.py respectively.

Offline data collection

1. Run python/RL/train.py to start the RL agent server.
2. Run the Java class job.TraSimSearch/TraRangeQuery/TraCluSearch to collect the offline data.
3. Run python/RL/offlineData.py to process the offline data.

Online process

1. Run python/RL/test.py to start the RL agent server.
2. Run the Java class job.TraSimSearch/TraRangeQuery/TraCluSearch to start the online process.

Datasets

We use two real-world datasets: T-Drive, Chengdu Taxi and one synthetic dataset: BrinkHoff.