Computerized dietary meal planning can support the prevention or management of chronic diseases. We propose a simple and powerful framework based on hybrid evolutionary computation to generate 10-day meal plans observing various requirements. The method is based on ontologies for modeling expert concepts, a structural model, and a bonus/malus rule base with numerical, harmony, and occurrence rules, arranged in user-defined aspects. For the search, we use NSGA-II combined with a reinforcement-learning-based mutation operator. The strengths of the aspects and rules are assigned by users or may also be tuned automatically. We implemented the ontology and the rule base using a hospital food database of a realistic size (185 food items), some important common-sense culinary practices, and the official dietary guidelines for public catering in Hungary (31 rules using 73 ontology sets). For validation, we compared 10-day plans to three alternative methods and also manually created plans. Our framework significantly outperformed the alternatives and the human expert in fitness and scores. However, a subjective evaluation by another expert found human-made plans to be superior to generated plans, especially in taste harmony due to the common-sense knowledge missing from the rule base. The results demonstrate that the framework is an effective tool for multi-objective dietary planning and can successfully balance competing requirements.
Alharan et al. (Tue,) studied this question.