Local Search over Relational Databases

Solving combinatorial problems is increasingly crucial in business applications, in order to cope with hard problems of practical relevance. However, the approach of exploiting, in production scenarios, current constraint or mathematical programming solvers has severe limitations, which demand new methods: data is usually stored in potentially large relational databases, and maintaining the problem in central memory, as required by current solvers, could be expensive, challenging, or even impossible, due to the size of the data and the possibly unacceptable loss of data integrity. We present a declarative language based on sql for modelling combinatorial problems, and novel techniques for local search algorithms explicitly designed to work directly on relational databases, also addressing the different cost model of querying data in the new framework. We also discuss and experiment with a solver implementation that, working on top of any relational DBMS, exploits such algorithms in a way transparent to the user, making a step forward to the seamless integration of combinatorial problem solving into business environments.