We present two dual control approaches to the model maintenance problem based on adaptive model predictive control (MPC). The controllers employ systematic self-excitation and design experiments that are performed under normal operation, resulting in improved control performance with smaller output variance and less control effort. Our control formulations offer a novel approach to the question of how to excite the plant input to generate informative data within the context of MPC and adaptive control. One controller actively tries to reduce the parameter-estimate error covariances; the other controller maximizes the information in the signals for enhanced learning. Our approach differs from existing ones in that we let our controllers converge to standard certainty equivalence (a) MPC when the parameter uncertainty decreases or more information is generated, and as a result we avoid plant excitation when the uncertainty is low or enough information has been generated. We demonstrate that the controllers work well with a large number of tuning configurations and also address the issue of models that are not admissible for control design. (C) 2015 Elsevier Ltd. All rights reserved.