A large number of websites use online recommendations to make web users interested in their products or content. Since no single recommendation approach is always best it is necessary to effectively combine different recommendation algorithms. This paper describes the architecture of a rule-based recommendation system which combines recommendations from different algorithms in a single recommendation database. Reinforcement learning is applied to continuously evaluate the users' acceptance of presented recommendations and to adapt the recommendations to reflect the users' interests. We describe the general architecture of the system, the database structure, the learning algorithm and the test setting for assessing the quality of the approach.