In the capacitated clustering problem (CCP), a given set of n weighted points is to be partitioned into p clusters such that the total weight of the points in each cluster does not exceed a given cluster capacity. The objective is to find a set of p centers that minimises the total scatter of points allocated to these centers. In this paper, we propose a merger of Greedy Random Adaptive Search Procedure (GRASP) and Adaptive Memory Programming (AMP) into a new GRAMPS framework for the CCP. A learning process is kept in charge of tracking information on the best components in an elite set of GRAMPS solutions. The information is strategically combined with problem-domain data to restart the construction search phase. At early stage of constructions, priorities are given to problem-domain data and progressively shifted towards generated information as the learning increases. GRAMPS is implemented with an efficient local search descent based on a restricted λ-interchange neighbourhood. Extensive experiments are reported on a standard set of bench-marks from the literature and on a new set of large instances. The results show that GRAMPS has an efficient learning mechanism and is competitive with the existing methods in the literature.
