Combinatorial optimization problems, a famous example of which is the Traveling Salesman Problem, are often solved with hand-crafted algorithms that are often costly to develop and difficult generalize. With this project, we investigate how state-of-the-art models can be trained to learn such heuristics automatically via reinforcement learning, and propose additional tweaks to improve its performance.
-
Furkan Gürsoy - furkangursoy.github.io
-
Batuhan Koyuncu - bkoyuncu.github.io
*Both contributed equally.
**We also thank Sercan Amaç for his contributions.
-
COwRL.pdf file contains the slides describing the problem and the performed analysis.
-
To train models using different distributions, you can replace the content of attention-learn-to-route/problems/tsp/problem_tsp.py file with the contents of problem_tsp_normal.py and problem_tsp_exp.py for normal and exponential distributions respectively.
-
pretrained_novel folder contains models trained with data from different distributions rather than the uniform distribution utilied in the original work.
-
TSP20_Experiments.ipynb and TSP50_Experiments.ipynb are two standalone notebooks for evaluating performance results of existing and novel pretrained models, and baselines on data from different distributions.
Feel free to open issues or contact us.