Reference point based multi-objective evolutionary algorithm by decomposition (RMEAD) considers reference points as users' preferences. RMEAD not only focuses on searching the region of interest to find a set of preferred solutions, but also economizes a significant amount of computing resources. However, the base weight vectors in RMEAD may not be well estimated when confronting to hard multi-objective optimization problems. This paper modifies RMEAD to improve its performance on two aspects: firstly, a novel and simple approach to finding the base weight vectors is developed, the correctness of which is proved mathematically; secondly, a new updating weight vectors method is proposed. Abundant experiments show that the improved RMEAD (IRMEAD) could obtain significantly better results than RMEAD on all the test cases in terms of convergence and diversity. Besides, compared with recent proposed preference-based approach MOEA/D-PRE, IRMEAD outperforms it on most of the test instances.