Safety-Driven Deep Reinforcement Learning for Cooperative Highway Ramp Merging

Description:

1. Designed and trained a deep reinforcement learning model to perform single-agent motion planning for vehicles involved in a highway ramp merging scenario using Ray and SUMO (Simulation of Urban MObility), based on an open source project called FLOW.

2. Designed and implemented safe-driven features for the model by applying an intelligent ramp meter in the merging scenario.

Flow

Flow is a computational framework for deep RL and control experiments for traffic microsimulation.

Software Setup:

1. Go through the Flow Installation Instructions to install Flow, SUMO and RLlib.
2. Test the installation according to the same instruction listed above.

Training:

1. Go to the flow/example/ folder and run: conda activate flow
2. For training the original highway merging scenario, run python train.py singleagent_merge --rl_trainer "rllib"
3. For training the modified highway merging scenario with ramp meter, run python train.py singleagent_ramp_meter --rl_trainer "rllib"
4. The trained results will be stored in ray_result which is outside the flow directory. It is stored in the form of checkpoints.

Visualization for original highway merging scenario:

1. Selet the folder inside the ray_result mentioned above, e.g. ray_results/stabilizing_open_network_merges/merge_1, note: merge_1 is an arbitrary name here.
2. Within the selected folder, select the checkpoint folder that you want to visualize, e.g. ray_results/stabilizing_open_network_merges/ramp_meter_1/500, where 500 is the number of the selected checkpoint.
3. Go to the flow/flow/visualize/ folder and run: conda activate flow
4. Given the example information colleceted above inside the ray_result folder, run python ./visualizer_rllib.py ~/ray_results/stabilizing_open_network_merges/merge_1 500, where merge_1 is the folder name selected above, and the 500 is the number of the selected checkpoint.

Visualization for modified highway merging with ramp meter scenario:

1. Selet the folder inside the ray_result mentioned above, e.g. ray_results/ramp_meter/merge_2, note: merge_2 is an arbitrary name here.
2. Within the selected folder, select the checkpoint folder that you want to visualize, e.g. ray_results/ramp_meter/ramp_meter_1/700, where 700 is the number of the selected checkpoint.
3. Go to the flow/flow/visualize/ folder and run: conda activate flow
4. Given the example information colleceted above inside the ray_result folder, run python ./visualizer_rllib.py ~/ray_results/ramp_meter/merge_2 700, where merge_2 is the folder name selected above, and the 700 is the number of the selected checkpoint.

Modifications from the original Flow (how I created the new highway merging with ramp meter scenario):

1. Created flow/examples/exp_configs/rl/singleagent/singleagent_ramp_meter.py, which indicates the training parameters, vehicle parameters, and flow parameters.
2. Created flow/flow/envs/ramp_meter.py, which determines the Reinforcement Learning functions including States, Actions, Rewards, etc.
3. Added the 'RampMeterPOEnv' into the fileflow/flow/envs/__init__.py, under the __all__ = [ session, so that the system can have access to the file.
4. Created flow/flow/networks/ramp_meter.py, which determines the SUMO traffic network as the training environment.
5. Added the from flow.networks.ramp_meter import RampMeterNetwork into the fileflow/flow/networks/__init__.py so that the system can have access to the file.