On probabilistic-method-based ML4CO methods

[bokveizen]

Dear authors,

Thanks a lot for your insightful work. I really enjoy reading it.

I am doing research in ML4CO, and have been focusing on probabilistic-method-based methods. Roughly speaking, we understand the output of neural networks as a solution distribution and try to directly compute the expected objective on the output distribution. Due to their directly interpretable nature, probabilistic-method-based methods have theoretical bounds for the final solution quality.

At the same time, I largely agree with your conclusion in your paper that the current ML4CO methods still need post-processing to obtain good final solutions. In the probabilistic-method-based case, from simple to sophisticated, we have been using the following post-processing methods: (1) direct sampling from the learned distribution, (2) entry-wise derandomization, and (3) greedy derandomization. Compared to MCTS, I believe the above three post-processing methods are still relatively simple.

What's your opinion on that? Do you think that would be a promising direction to further explore?

Feel free to check my recent work and the references thereof :D

• https://arxiv.org/abs/2405.08424
• https://arxiv.org/abs/2207.05984
• https://arxiv.org/abs/2006.10643

Best, Fanchen