A revised machine learning methodology is applied to a simulated crystallization process flowsheet for continual improvement of its performance by generating and analyzing process data. The aim is to identify bands of crucial decision variables leading to zones of best average process performance. The methodology comprises two components : symbolic induction and case-based reasoning. It uses an incremental algorithm to update performance classification rules, which not only improves the efficiency of the symbolic induction of the classification rules by eliminating the need for their periodic reinduction but also simplifies the case-based reasoning step. The new concepts and procedures are illustrated by application to a potassium nitrate crystallization process comprising a mixed-suspension, mixed-product-removal crystallizer, a hydrocyclone, and a fines dissolver. By identifying and establishing ranges for the most crucial decision variables, namely feed concentration, flow rate, and cooling stream temperature, three zones leaning to an improvement of nearly 12% on nominal average performance are detected within two generations of the classification rules.