Reinforcement Learning Final Project Code
- Python 3.7
- pytorch 1.3.1
- matplotlib
- pandas
For dynamic TSP, nearest neighbor is used as the baseline. To run the benchmark, go to dynamic_tsp/dynamic_nn and run the command python dynalocTSP.py.
The DRL code for dynamic TSP is in dynamic_tsp/drl.
To train the model, run python trainer.py.
To run the benchmark using the provided checkpoint for {num_nodes} number of nodes, run python trainer.py --checkpoint tsp/20/12_46_22.667847/ --test --batch_size 1 ---nodes {num_nodes}.
For dynamic VRP, a variation of the Clarke-Wright algorithm is used as the baseline. This variation generates routes like normal, but immediately returns to the depot if the current route would violate demand constraints due to the dynamic demand. The algorithm is rerun (with visited nodes ignored) each time this happens. To run the benchmark, go to dynamic_drp/dynamic_cw and run the command python dynademandVRP.py.
The DRL code for dynamic VRP is in dynamic_vrp/drl.
To train the model, run python trainer.py --task vrp.
To run the benchmark using the provided checkpoint for {num_nodes} number of nodes, run python trainer.py --checkpoint vrp/20/22_55_44.077690/ --test --batch_size 1 ---nodes {num_nodes}.