TSPTW as a great benchmark tool.

[felixchalumeau]

It would be great to make it possible to work on the TSPTW as it is one of the problem used by Quentin in his paper.

In addition, even it could be a little time consuming, I think it could be very interesting to try to reproduce his experiment with our solver. Not a perfect reproduction as we are learning inside the solver, but using the same Graph representation as him.

Even if this won't be inside the "unified representation goal" we are trying to achieve, I would see it as a "sanity check", to make sure our performance are good and to prove that learning in the solver is a real breakthrough.

To do it, we should:

• create a tsptw generator (similar as Quentin's one)
• resolve Make it possible to change easily the way RL represention is made #21 first (making sure we can easily change the RL representation outside the package).
• make sure we can use similar architecture as his (not necessarly exactly the same though)

PS: the comparison with Quentin can also be done with knapsack or portfolio but in the end, I think we should have the tsptw in our examples

[ilancoulon]

It would definitely be great to do that, but I'm not sure we have time for this.

[felixchalumeau]

If it is great, we should find the time ! ;)
Let's say it is not a priority but I think it would be a great element to add to our paper !

[ilancoulon]

We are going to try that.

Here would be the list of the things we need to develop:

• BoolVar, done in Adding BoolVar & BoolDomain #54
• ReifiedLessOrEqual constraint, cc @felixchalumeau Add Islessorequal constraint (reified less or equal constraint) #57
• Or constraint, cc @felixchalumeau
• IntSetVar, cc @ilancoulon done in Add IntVarSet #58
• Element2DConstraint, done in Element2d & BinaryMaximum (bound consistent) constraint #67
• MaxConstraint -> BinaryMaximum constraint, done in Element2d & BinaryMaximum (bound consistent) constraint #67
• IntInSetConstraint cc @ilancoulon, done in InSet & ReifiedInSet #60
• ReifiedIntInSet cc @ilancoulon , done in InSet & ReifiedInSet #60
• A = B\x (with A, B being sets, and x an int) cc @ilancoulon , done Add SetDiffSingleton constraint #66
• BoolVarViewNegation || Negation constraint done in SetEqualConstant constraint & BoolVarViewNot #68
• SetEqualConstant constraint, done in SetEqualConstant constraint & BoolVarViewNot #68
• Element1D constraint, done in Element1D constraint #69
• TsptwGenerator, done in Tsptw generator #70
• TsptwStateRepresentation, done in Tsptw staterepresentation #73
• TsptwReward, done in Refactor Reward and add TsptwReward #75
• Add a out_solver parameter to the search function, done ine New feature : training like Quentin, "out of solver" #74
• Reproduce Quentin's experience (as close as possible) : sanity check of the solver
• Produce a new experience taking advantage of the solver (e.g. improving reward engineering) : prove of concept of the solver.

Eventually :

• Prove the interest further by trying a double-agent (solution oriented with optimality oriented)

[ilancoulon]

We could consider it done imo :)