3 ms·

In general, modern constraint programming solvers runs the propagators for the constraints until fixpoint. The level of propagation for a constraint is typicall

• dragontamer 5y ago
  Isn't the "benchmark" for typical propagators arc consistency? I'm not too familiar with many kinds of constraints. But the constraints I am familiar with such as "Global Cardinality Constraint" (which is a generalization upon all_different) reaches arc-consistency after a basic "maximal flow" algorithm is applied to it. Indeed: proving that the maximum-flow algorithm over global-cardinality-constraint (aka all_different) reaches arc-consistency proves that you can reach arc-consistency over the whole problem set by iterating over the fixed point. ------- In effect, we quickly learn to reach arc-consistency using a different algorithm (ex: maximum flow) over the "all_different(x1, …, xn)". You're right in that "arc-consistency classic" is not being done here. But the effect is the same. Maximum-flow(all_different(x1, x2... xn)) will lead to arc-consistency with the all_different constraint. Now that you're "locally arc-consistent", you iterate over "a+b<=c" and make that constraint arc-consistent. If the domains are restricted again, you run the maximum-flow algorithm over all_different(x1, x2... xn) and achieve arc-consistency again. (Bonus points: the "last time" you ran maximum_flow is still relevant and serves to accelerate the 2nd run of the maximum_flow problem. That residual graph is still correct!! You just need to slightly update it given the new restrictions on the domains) ------- In case you're curious: the all_different constraint being arc-consistent with maximum flow is pretty fun to think about. All variables are on one-side of a bipartite graph, while all assignments are on the 2nd side of a bipartite graph. Connect all variables to assigments that match their domains. Connect a "super-source" to all variables, and "super-sink" to all values. Super-sink has a value of "1" for each connection to each variable (aka: each variable can only be assigned once). Solve for maximum flow from super-source to the super-sink. Now that you have one particular maximum-flow (a possible assignment), it is possible to determine all other assignments by iterating the possibilities where the flow goes but remains maximum.
  • mzl 5y ago
    For many constraints there are propagators that can reach domain consistency (all remaining values for all variables are part of some solution to the constraint, sometimes called generalized arc consistency). For some constraints the best propagator in practice is a much weaker one, as propagating to domain consistency is in itself NP-hard. Many times there is no actual formal consistency level definable. A great example of this is the circuit constraint that holds if a set of variables forms a Hamiltonian path in a graph. Hamiltonian paths is a classic NP-complete problem. The propagators for it make smart inferences, but not as a consistency level. Crucially, propagators are for a single constraint only. For a constraint problem, you run all the propagators to a fix point. There is no single global “algorithm” here, and there are no higher level consistencies.
    • dragontamer 5y ago
      But "Arc Consistency" doesn't solve a problem. It only removes "obviously wrong" assigments from a variable's domain. The arc-consistency to the circuit constraint is NOT about solving Hamiltonian paths. Its about finding all "obviously wrong" assignments that can be discovered looking at 2-variables at a time. Because all arc-consistency algorithms can be brute-forced by looking at all combinations of pairs of variables, we have at worst O(n^2) number of variables to consider to achieve arc-consistency. Each variable only has a finite number of assignments (ignoring ILP with infinite numbers: assuming classical constraint programming). Which means that you can brute-force all possibilities in O(n^2 * d^2) where d is the cardinality of the domain, and n is the number of variables. This means that reaching arc-consistency over the circuit / Hamiltonian path problem is in O(n^2*d^2) worst case (per propagation step. Another O(n) multiplier is added for reaching arc consistency over all variables), with possibly a specialized algorithm that goes faster than that brute force style.
      • mzl 5y ago
        Arc consistency is not really a good term to use, since it was originally defined for 2-ary constraints only, but is sometimes used for n-ary constraints. That is why I use the term domain consistency consistently :-) Given a propagator for circuit that reaches domain consistency (note, for the whole n-ary constraint), that is equivalent to a checker of there is a Hamiltonian path and that will most likely never be a polynomial algorithm (unless P=NP). I can’t say that I’m interested in any particular property of the decomposition of the constraint into 2-ary parts, it is not that interesting of a structure. The cool filterings that circuit propagators do are based on global analysis of the whole graph, not on any local property. For example by finding ordering a of strongly connected components in the current graph. Almost all propagators need to be reasonably fast to be usable, that is sort of a given in this area. An exponential propagator is almost always a bad idea. For all_different there is a O(n^2.5) algorithm for domain consistency which is really cool and important. In practice, the simple value propagation (remove assigned values from the other variables) is often better for finding solutions quickly.