With the increase of online services and the development of e-commerce, there are more and more complex routing problems to solve. However such problems are generally hard to solve optimally because they are generalisations of the TSP which is NP-Hard. Therefore many approximation algorithms such as metaheuristics have been designed to compute near optimal solutions with high accuracy for these problems. Sometimes the problem needs to optimize many functions at the same time and hence requires a multi-objective approach to solve it. If some metaheuristics are already used for few decades, new trends try to integrate machine learning methods into existing metaheuristics to improve their performance. In fact the machine learning phase tries to discover some knowledge about the problem and then helps to efficiently cross the solution space. If such techniques are commonly used in single-objective problems, they can not be applied directly for multi-objective problems. This article presents a new learning mechanism which helps solving multi-objective routing problems with time windows.