Machine learned heuristics to improve constraint satisfaction

Although propagation techniques are very important to solve constraint solving problems, heuristics are still necessary to handle non trivial problems efficiently. General principles may be defined for such heuristics (e.g. first-fail and best-promise), but problems arise in their implementation except for some limited sources of information (e.g. cardinality of variables domain). Other possibly relevant features are ignored due to the difficulty in understanding their interaction and a convenient way of integrating them. In this paper we illustrate such difficulties in a specific problem, determination of protein structure from Nuclear Magnetic Resonance (NMR) data. We show that machine learning techniques can be used to define better heuristics than the use of heuristics based on single features, or even than their combination in simple form (e.g majority vote). The technique is quite general and, with the necessary adaptations, may be applied to many other constraint satisfaction problems.