Diet models with linear Goal Programming: impact of achievement functions 10 February 2016 1 Appendix A. Model Class 2b: Minimize the differences between the actual diet and the proposed diet 2 3 Models in Class 2b search for diets that conform as closely as possible to an actual diet, while 4 meeting constraints on e.g. nutrient recommendations, palatability and total cost. 5 The composition of the actual and proposed diet can be expressed in amounts consumed per 6 food group1;2 or in energy provided per food group3. These result in models of similar 7 structure, and therefore we only elaborate models that express diets in amounts consumed per 8 food group. 9 10 Notation, GP constraint 11 In this appendix we use the following notation: 12 13 Qi ~ Data: Intake of food group i in actual diet (i = 1…I) 14 Xi ~ Decision variable: Intake of food group i in proposed diet (i = 1…I) 15 16 with I the number of food groups. The deviation between the actual diet and the proposed diet 17 can be calculated via a GP constraint: 18 19 X i Qi d i Qi d i Qi 20 d i , d i 0 for all food groups i 21 22 Now di di di is the absolute value of the relative deviation between the actual and the 23 proposed diet*. The negative and positive deviations d i , d i are normalized: di 1 means 24 that the consumption in food group i of the proposed diet is 100% lower than in the actual 25 diet, and di 1 means that the consumption in food group i of the proposed diet is 100% 26 higher than in the actual diet. 27 28 MinSum achievement function 29 As an example we consider a diet problem with two foods. Actual food intakes are Q1 = 3 and 30 Q2 = 2, and a nutritional constraint 2X1 + 3X2 ≥ 15 applies. The proposed diet should resemble 31 the actual diet as much as possible. Therefore, all deviations d i , d i from the actual diet are 32 regarded as unwanted. A MinSum GP model could look as follows: Note that it is also possible to model absolute deviations: X i d i d i Qi for all i. One should keep in mind that any choice made in modeling the deviation uses judgment from the decision-maker and implies assumptions on the underlying preference structure. * Diet models with linear Goal Programming: impact of achievement functions 10 February 2016 33 34 Minimize D sum X1 35 2 X1 36 X2 3X 2 w1 d1 w1 d1 w2 d 2 w2 d 2 3d1 2d 2 3d1 2d 2 [A1] 3 [A 2] 2 [A3] 15 [A 4] X 1 , X 2 , Dsum , d1 , d1 , d 2 , d 2 , d3 , d3 0 37 with w1 , w1 , w2 , w2 the non-negative weights assigned by the user. 38 The shaded area in Figure A1 contains all diets that comply with nutritional constraint [A4]. 39 Of these we want to find the one that has minimal deviation from the actual diet. The set of 40 Pareto-optimal solutions is indicated with a bold line segment connecting (3,3) and (4.5,2). 41 42 43 44 45 46 47 Figure A1. Class 2b: The shaded area contains all diets that comply with nutritional constraint [A4]. Of these we aim to find the diet with minimal deviation from the actual diet ( ). A MinSum GP model will yield diet (4.5, 2) or diet (3,3) ( ). Diet (3.75, 2.5) ( ) is reachable via a MinMax GP model. The optimal solution of model [A1]-[A4] depends upon the weights that are chosen. However, 48 independent of the weights, it is always located within the Pareto set (i.e. the set of Pareto- 49 optimal solutions). Note that all Pareto-optimal solutions have d1 d 2 0 . For w1 w2 1 50 all Pareto-optimal solutions have total weighted deviation Dsum = 1. That means that – 51 according to the MinSum GP model with w1 w2 1 – all Pareto-optimal diets are equally 52 preferable. Due to the fact that in LP-problems the optimal solution is always found in a 53 corner-point, an LP-solver will generate one of the corner-points of the efficient set: (4.5,2) or 54 (3,3). However, if the user specifies that w1 w2 (s)he expresses that (s)he cannot justify 55 assigning one deviation more importance than another, and therefore it would be natural to 56 obtain a balanced solution that spreads the deviations over both deviational variables Diet models with linear Goal Programming: impact of achievement functions 10 February 2016 57 ( d1 , d 2 ) instead of an unbalanced solution that piles the unwanted deviation on only one of 58 them. 59 60 61 62 63 64 65 Table A1. Optimal solution of MinSum GP model [A1]-[A4] for various choices of the weights. Optimal solution Weights Diet Deviations Weighted deviations Achievement function ( w1 , w2 ) ( X1 , X 2 ) ( d1 , d 2 ) ( w1 d1 , w2 d 2 ) (0.9, 1) (4.5, 2) (0.5, 0) (0.45, 0) 0.45 (1, 1) (4.5, 2) or (3, 3)† (0.5, 0) or (0, 0.5) (0.45, 0) or (0, 0.45) 0.45 (1, 0.9) (3, 3) (0, 0.5) (0, 0.45) 0.45 Dsum = w1 d1 w2 d 2 † According to the MinSum achievement function all diets on line segment (4.5, 2) – (3, 3) are equally preferable, because they all have the same achievement function value: 0.45. However, an LP-solver will only generate a corner-point: (4.5, 2) or (3, 3). Table A1 shows the optimal solution of model [A1]-[A4] for various choices of the weights 66 w1 , w2 . It shows that the MinSum model is sensitive to weight changes: slight changes in the 67 weights make the optimal diet ‘jump’ from one corner-point to another. 68 69 MinMax achievement function 70 The MinMax GP version of model [A1]-[A4] is 71 72 Minimize { Dmax } [A5] X1 73 2 X1 74 75 X2 3X 2 3d1 2d 2 3d1 2d 2 3 [A 2] 2 [A3] 15 [A 4] Dmax w1 d1 Dmax w1 d1 [A6] [A7] Dmax w2 d 2 Dmax w2 d 2 [A8] [A9] X1, X2 ≥ 0; Dmax , d1 , d1 , d 2 , d 2 0 76 77 Table A2 shows the optimal solution of model [A2]-[A9] for various choices of the weights 78 w1 , w2 . Now, for w1 w2 1 a balanced solution is obtained, in which all unwanted 79 deviations are equal. Moreover, the model is less sensitive to weight changes: slight changes 80 in the weights cause minor shifts in the optimal solution. The MinMax GP model represents a 81 situation in which the resemblance between two diets is determined by the food that differs 82 the most. Diet models with linear Goal Programming: impact of achievement functions 10 February 2016 83 84 Table A2. Optimal solution of model MinMax GP model [A2]-[A9] for various choices of the weights. Optimal solution Weights Diet Deviations Weighted deviations ( w1 , w2 ) ( X1 , X 2 ) ( d1 , d 2 ) ( w1 d1 , w2 d 2 ) Achievement function Dmax (0.9, 1) (3.79, 2.47) (0.263, 0.237) (0.237, 0.237) 0.237 (1, 1) (3.75, 2.50) (0.250, 0.250) (0.250, 0.250) 0.250 (1, 0.9) (3.71, 2.53) (0.237, 0.263) (0.237, 0.237) 0.237 85 86 Extended GP achievement function 87 The achievement functions of the MinSum and MinMax GP model can be combined into the 88 achievement function Dext of a so-called Extended GP model 4;5: 89 90 Minimize Dext (1 ) Dsum Dmax [A10] 91 92 where parameter 0;1 weighs the importance attached to the minimization of the MinSum 93 and MinMax achievement function. An Extended GP model can find solutions that are 94 compromises between the MinSum and MinMax solutions. It can thus provide valuable 95 dietary suggestions that are located ‘between’ the MinSum and MinMax diets. However, the 96 numerical example is too small to contain any of these intermediate solutions. 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 References 1. Maillot M, Vieux F, Amiot MJ, Darmon N (2010) Individual diet modeling translates nutrient recommendations into realistic and individual-specific food choices. American Journal of Clinical Nutrition 91, 421-430. 2. Masset G, Monsivais P, Maillot M, Darmon N, Drewnowski A (2009) Diet optimization methods can help translate dietary guidelines into a cancer prevention food plan. Journal of Nutrition 139, 1541-1548. 3. Darmon N, Ferguson EL, Briend A (2006) Impact of a Cost Constraint on Nutritionally Adequate Food Choices for French Women: An Analysis by Linear Programming. Journal of Nutrition Education and Behavior 38, 82-90. 4. Romero C (2004) A general structure of achievement function for a goal programming model. European Journal of Operational Research 153, 675-686. 5. Romero C, Tamiz M, Jones DF (1998) Goal Programming, Compromise Programming and Reference Point Method Formulations: Linkages and Utility Interpretations. The Journal of the Operational Research Society 49, 986-991.