A Family of Stochastic Methods For Constraint Satisfaction and Optimisation

Constraint satisfaction and optimisation is NP-complete by nature. The combinatorial explosion problem prevents complete constraint programming methods from solving many real-life constraint problems. In many situations, stochastic search methods, many of which sacrifice completeness for efficiency, are needed. This paper reports a family of stochastic algorithms for constraint satisfaction and optimisation. Developed with hardware implementation in mind, GENET is a class of computation models for constraint satisfaction. Genet is a connectionist approach. A problem is represented by a network with inhibitory connections. The network is designed to converge, in a fashion that resembles the min-conflict repair method. Reinforcement learning is used to bring GENET out of local optima. Building upon GENET as well as ideas from operations research, Guided Local Search (GLS) and Fast Local Search are novel meta-heuristic search methods for constraint optimisation. GLS sits on top of other local-search algorithms. The basic principle of GLS is to penalise features exhibited by the candidate solution when a local search settles in a local optimum. FLS is a way of reducing the size of the neighbourhood so as to improve the efficiency of local search. As a meta-heuristic, GLS is embedded in genetic algorithms to form the Guided Genetic Algorithm (GGA). GGA extends the domain of application by GLS and improves its reliability (i.e. getting good results consistently). GENET, GLS, FLS and GGA have been applied to a non-trivial number of satisfiability and optimisation problems and achieved world-class results. GLS has also been incorporated in ILOG Dispatcher, a commercial package for vehicle routing.