Multi-criteria decisions taken by autonomous systems cannot rely on external decision makers (DM) to work properly. We propose a new multicriteria decision making method suitable for autonomous systems, which uses a set of nested quality boxes (like Russian dolls) in the criteria space, to define an utility function. A learning method is proposed to configure the boxes from subjective quality assessments provided by the autonomous system users. We have tested the performance of the Russian doll-like MCDM method to multi-criteria route selection for maximizing the Quality of Service in wireless ad hoc networks. Experiments have shown that route selection controlled by the same Russian doll structure for several scenarios is able to achieve the best user satisfactions for all test cases. Conversely, the weighted sum method has required to specifically adjust its weights according to each scenario to reach the best satisfaction scores.