The study introduces an innovative enhancement of the existing single-objective Q-learning induced bee colony (QLNBC) algorithm in multi-objective setting to solve noisy multi-objective optimization (MOO) problem. The introduced algorithm, referred to as multi-population multi-objective QLNBC (MPMO-QLNBC) encompasses four extensions. First, to solve a noisy MOO problem with N objectives, N QLNBC algorithms are run in parallel for optimization of individual objectives. Each QLNBC employs a temporal difference Q-learning (TDQL) to guide a trial solution of the concerned population to select a sample size for periodic evaluation of a specific objective only. This strategy helps in exhaustively exploiting each fitness terrain to get rid of intrusion of noise in the fitness measures of trial solutions. Second, a controlled migration policy is devised to allow communication among the members of the multiple QLNBC populations. It assists in employing an effective way of knowledge transfer across multiple fitness landscapes and enhancing population diversity, especially in the predominant presence of noise in the multi-modal objective surfaces. Third, a strategy is proposed to select a group of solutions, instead of only the best solution, from each population which reduces the risk of discarding a good solution. Finally, a unique ranking policy is designed to find the set of equally good solutions of the noisy MOO problem. Tests were done using noisy versions of 23 benchmark functions from CEC 2009, and the results showed that the suggested MPMO-QLNBC performed noticeably better in terms of quality compared to other methods.