Abstract : In this PhD thesis, we study several tree search methods for solving CSPs and focus on the discrepancy concept (a discrepancy is a deviation from the first choice of the heuristic). In this context, we propose improving mechanisms for general methods. These mechanisms take benefits from conflicts and guide the search by weighting the variables and the values. We propose also special mechanisms for methods based on discrepancies as the discrepancies restriction, the discrepancies counting, and the discrepancies positions. All propostions are validated by experiments done on real and random CSPs. We compare variants of methods based on discrepancies integrating several combinations of imrprovements and other methods known for their efficiency. In a second part, we extend our propositions to an optimisation context considering scheduling problems with time lags. In this purpose, we adapt a discrepancy-based method, Climbing Discrepancy Search, to solve these problems. Efficiency of some improved variants of this method is tested on known benchmarks.