Adaptive Strategies for Solving Constraint Satisfaction Problems

A major challenge in constraint programming is to develop efficient generic approaches to solve instances of the constraint satisfaction problem (CSP). In recent years, adaptive approaches for solving CSPs have attracted the interest of many researchers. General speaking, a strategy that uses the results of its own search experience to modify its subsequent behavior does adaptive search. In this dissertationwe explore adaptive strategies for backtracking search on various levels. First, we investigate adaptive search-guiding heuristics for ordering variables in CSPs. These adaptive heuristics learn and use information from every node explored in the search tree, whereas traditional static and dynamic heuristics only use information about the initial and current nodes. We then perform a wide empirical evaluation of the proposed variable ordering heuristics and compare them with the current state-of-the-art variable ordering strategies. Concerning constraint propagationwhich is used as an inferencemechanism in order to simplify a problem so as to make it easier to solve, we explore adaptive strategies for ordering the different revisions performed when enforcing arc consistency algorithms. Next, we propose adaptive branching heuristics for splitting the search tree. The application of these heuristics results in an adaptive branching scheme. Experiments with instantiations of the proposed generic heuristics confirm that search with adaptive branching outperforms search with a fixed branching scheme on a wide range of problem. Finally, we propose a new a generic approach for branching where the variable’s domains are grouped into sets by using the scores assigned to values by a value ordering heuristic, and a clustering algorithm from machine learning. In general, this dissertation contributes to the design and implementation of adaptive and autonomous constraint solvers that have the ability to advantageously modify modelers decisions that typically in mainstream CP solvers are taken prior to search.