Dietary planning using multi objective evolutionary algorithmwith fuzzy preference integration

Nutrition is the process of breaking down and digestion of food in order to supply energy and other essential nutrients to the organism. Since eating is the main method to sustain life and survive, it has always been a forceful motivator for human-beings. It covers such an important part of human beings' daily activities that throughout the history they struggled to reach food easily in order to be healthy, powerful and fit. Archaeologist, anthropologist, sociologist and biologist who study history of human evolutionary process often used strategies to obtain food as a complementary tool. Early human beings are suggested to use scavenging and later on they started to hunt animals and gather fruits etc. to feed. After learning techniques related to agriculture, gardening and domestication of animals humans achieved to optimize resources such as time and energy to attain food. Throughout decades, people living in different physical conditions have evolved various food gathering methods and diets by using resources in their habitat. Although this optimization came in handy, changing diet brought some problems with it. Today modern world faces a serious malnutrition crisis. According to Global Nutrition Report 2 billion people lack micro-nutrients, 155 million children are stunted 42 million children are wasted, 2 billion adults are overweight or obese. This indicates clearly that a form of nutritional planning is required to overcome this problems. Dietary planning is an effort to develop guidelines on nutritional requirements to improve mental and physical fitness of people. Countries try to establish and pursue dietary planning policies for their citizens. In order to breed healthy populations a set of reference nutritional intake values must be devised and maintained in order to measure and improve quality of a diet. This measures are called dietary reference intake (DRI) and they reflect the state of scientific research on nutritional requirements. Although DRI measures imposes some constraints on dietary planning, there are also economic, social and personal perspectives to the problem. In this thesis we first model a many-objective dietary planning problem with 4 distinct objectives. First objective is to minimize the cost from an economic perspective. Second objective is to optimize rating objective which reflects society's preference on that diet. Third objective is to optimize on preference objective which reflect personal taste and last objective is to optimize on preparation time which represents resources spent to attain that diet. Since dietary planning is a highly personalized problem these objectives have different importance levels for different users. After modelling many-objective dietary planning problem, a solution using preference based multi-objective evolutionary algorithm namely NSGA-II is proposed. In order to integrate users attitude towards importance of each objectives other than cost a desirability function approach is taken. Desirability functions map objective space into desirability space using exponential functions. This approach is proven to be successfull in many different domains but seriously lacks expressiveness. As a novel contribution to the field a Mamdani type fuzzy inference system is developed to integrate users preferences over different objectives. In order to limit search space of many-objective diet problem FIS takes preference, rating and preparation time as inputs and produces a desirability measure using the FIS rule set that is given by the user herself/himself. This approach brings in serious advantages. First one is the flexibility of user preference integration. This enables to reuse same optimization algorithm without changing for optimizing different aspects of the diet problem by just changing rule sets of FIS. Another advantage is the usage of verbal rule-sets such as "if rating is high then desirability is optimal". This intuitive and powerful way of expressing preferences increases accessibility of the system from users perspective. Another advantage is the dimensional reduction which decreases the time to compute pareto fronts in the region of interest. With mapping many-objective diet problem into a bi-objective diet problem using FIS, it became possible to use NSGA-II to solve the problem. An indirect permutation representation containing 317 foods in test data set is used for its advantages in constraint handling. Output of NSGA-II is a pareto front where cost objective is minimized and desirability objective is maximized. Experiment results show that fuzzy inference systems can be used to guide MOEA algorithms to region of interest of decision maker in search space. Tests on NSGA-II parameters indicates that increase in iteration count increases the quality of solutions. Although minor differences are observed in terms of mutation operator and rate, this observations does not support a statistical difference. Experiments on FIS rule sets concludes statistical differences supporting our hypothesis on effectiveness of FIS in guiding MOEA. Proposed system can be used in personnel dietary planning and recommendation engines easily. This system can also be enriched with extra constraints related with different nutritional diseases such as diabetes to improve on their dietary planning in hospitals. Proposed fuzzy preference based MOEA can also be used with any multi objective optimization problems which contains imprecise, user dependent and vague objectives.