dimension of the problem to solve "real world" TSP/VRP
g-dendiev opened this issue · comments
Hello,
First, thank you for the code, it's very interesting and usefull, i've also read your paper and i'm interested in TSP and VRP problem.
I want to use this model but with a "real" dataset. The point is that if I want to respect all distances between my "n" cities, I have to project them into dimensions "n-1".
I've seen the next parameter in your Env class (both TSP and VRP) :
- args['input_dim']: dimension of the problem which is 2
I want to change it (it's set into task_specific_params.py) to solve "real world" problem.
Have you tried change this into "n-1" dimension (9 for TSP 10) to maintain all distances you have between each pair of the 10 nodes ?
I'm triing to use your code, to adapt it to "real" problems, in which you can not consider distances as the crow flies. The real agent need to follow roads.
I'm triing to work on such an issue, by generating new structure of datas to train and test in some new dimensions.
Maybe you also did ?
I've work on a paper and I know that it's dificult to adapt some code made to illustrate progress in a field to an industrial real world big problem.
I'll appreciate if you could tell me if you have been considering my type of approach.
Thanks again !
I have not personally tried to incorporate the distance from the other cities as the input features. I just use the distance as a notion of reward. If I understand correctly, you want to extend this work with the distances from the other nodes as an extra feature in order to improve the results, right? Otherwise, the current setting should work -- you only need to change the reward function.
I thought about adding distances as extra features, but I don't think that they will help a lot to improve the results. My intuition is that in addition to the distances, the relative position of the other nodes matter. For example, let's consider a node with 2 adjacent nodes. Depending on what is the distance of the other two nodes is, the problem structure changes, but only adding distance does not reflect too much about the structure of the problem. For more information, I refer you to the graph embedding literature e.g. GCN, etc.
Please let me know if I am missing your point. I am more than happy to discuss the details of your problem.
Dear g-dendiev, should I close this? Please let me know if you have any further questions.
--Reza
Hello g-dendiev,
Did you manage to implement the code with real world data (coordinates X and Y). I'm also trying to solve a "real world " routing problem with RL
thanks