We show that a straightforward extension of a simple learning model based on the Hebb rule, the previously introduced association-reinforcement Hebb rule, can cope with `delayed', unspecific reinforcement also in the case of structured data and lead to perfect generalization.