Humans frequently have to face complex problems. A classical approach to solve them is to search the solution by means of a trial and error method. This approach is often used with success by artificial systems. However, when facing highly complex problems, it becomes necessary to introduce control knowledge (heuristics) in order to limit the number of trials needed to find the optimal solution. Unfortunately, acquiring and maintaining such knowledge can be fastidious. In this paper, we propose an automatic knowledge revision approach for systems based on a trial and error method. Our approach allows to revise the knowledge off-line by means of experiments. It is based on the analysis of solved instances of the considered problem and on the exploration of the knowledge space. Indeed, we formulate the revision problem as a search problem: we search the knowledge set that maximises the performances of the system on a sample of problem instances. Our knowledge revision approach has been implemented for a real-world industrial application: automated cartographic generalisation, a complex task of the cartography domain. In this implementation, we demonstrate that our approach improves the quality of the knowledge and thus the performance of the system.