MARL-Traffic-System

// Below are the command lines needed to train and evaluate the code. We reccomend training the models in stages, with each stage having more randomization. For ex: the first stage is simple and just focuses on rewarding for driving, in the next stage it randomizes lane width, ect, and with stage 3 containing stages 2 randomness but with different acceleration,braking and more.

--env can be changed with the following "roundabout" "intersection" "tollgate"

STEP 1: activate your virtual env and run python train.py --stage 1 --env roundabout --num-agents 2 --workers 4 --train-batch-size 4000 --stop-iters 200

STEP 2: run //whatever the path you have to checkpoints python train.py --stage 2 --env roundabout --num-agents 4 --workers 6 --train-batch-size 8000 --stop-iters 300 --resume "C:\Code\MARL-Traffic-System\checkpoints"

Step 3: run python train.py --stage 3 --env roundabout --num-agents 6 --workers 6 --train-batch-size 10000 --stop-iters 400 --resume "C:\Code\MARL-Traffic-System\checkpoints"

Step 4: evaluate... if you want to test a different stage, simply chage stage, either 1,2,3 python eval.py --checkpoint "C:\Code\MARL-Traffic-System\checkpoints" --num-agents 2 --env roundabout --stage 3