Chapter 7 Nonlinear Optimization Models Sources of nonlinearity Non-constant returns Demand as a function of price is often nonlinear In investment decisions, risk is measured as a nonlinear function such as standard deviation NLP problem = Nonlinear programming problem Global optimum: Local optimum: A solution that is best in the entire feasible region A solution that is better than all nearby solutions in the feasible region Solver always finds global optimum solutions for LP problems. However, Solver often results in local optimum solutions for NLP problems. Convex function Slope is always decreasing y = cxa where a>0, c>0, x>0 y = cex where c > 0 Concave function Slope is always increasing y = cln(x) where c>0, and x>0 y = cxa where 0<a<1, c>0, x>0 Solver guarantees global optimum if: Objective function Conditions Minimization Objective function is convex Constraints are linear Maximization Objective function is concave Constraints are linear Pricing models Example 7.1 Pricing decision (Pricing1.xls) Suppose that the company estimates demand to be 400 when price is $70 and 300 if the price is $80. Estimate demand curve and determine the best price. Linear price curve: D = a + b x, where x = price (non-constant demand elasticity) Power price curve: D = axb, where x = price (constant demand elasticity) Sensitivity: Profit as price changes from $55 to $110 in increments of 5 Profit as elasticity changes from -2.4 to -1.8 in increments of .1 Example 7.2 Pricing decision with exchange rate (Pricing2.xls) Suppose that the demand is given by D = 27556760x-2.4, where x = price. The present exchange rate is $1 = £1.5599999. Determine the best price at which profit in $ is maximized. Sensitivity: What will happen to profit as the exchange rate fluctuates Example 7.3 Pricing suits at Sullivan’s (Pricing3.xls) Suppose that the current cost of suits is $320 and the corresponding demand is 300 suits per year. The elasticity of demand is constant at -2.5. Each purchase of a suit leads to an average purchase of 2 shirts and 1.5 ties. Profit margins are $25 for shirt and $15 for ties. Determine profit maximizing price. Sensitivity: What if suits are sold below cost? Also, what if the sales of secondary products shirts and ties are not the same as assumed? Example 7.4 Peak load pricing (Pricing4.xls) Demand is given by, Dp = 60 – 0.5Pp + 0.1Po, and Do = 40 – Po + 0.1Pp where, Dp = Peak demand in kwh, Pp = Peak price, per kwh Do = off-peak demand in kwh, and Po = off-peak price per kwh. Also, cost to main 1 kwh of power per day is $10. Determine pricing and capacity to maximize profit. Sensitivity: Change unit cost of capacity from $5 to $15 in increments of 1 Advertising response and selection models Example 7.5 Estimating an advertisement response function (AdvertisingResponse.xls) Given: Number of ads Exposures (millions) 1 4.7 8 22.1 20 48.7 50 90.3 100 130.5 Purpose: Estimate a response function that will estimate the number of exposures given number of ads. Plot of the given data reveals a nonlinear function which increases at a decreasing rate. There are many mathematical functions that have this basic shape. One such function is, f(n) = a(1 – e-bn) where n = number of ads and a and b are parameters to be found. Model Changing cells: a and b Target cell: Minimize RSME (Root mean square error) Constraints: None Example 7.6 Advertising selection (AdvertisingSelection.xls) Given: Advertising exposure = f(n) = a(1 – e-bn) where n = number of ads and a and b are parameters to be found. Table for values of parameter “a” Men 18-35 Men 36-55 Men >55 Women 18-35 Women 36-55 Women >55 Friends 93.061 61.129 33.376 105.803 71.784 56.828 Malcolm in Middle 84.772 61.528 9.913 66.998 46.146 8.887 MNF 116.808 76.527 57.84 40.113 26.534 17.209 Sports Center 43.647 47.749 30.075 22.101 16.151 9.101 TRL Live 26.711 19.655 10.751 42.451 34.609 8.46 Lifetime movie 11.99 10.281 11.51 29.403 24.276 31.149 CNN 11.793 9.982 22.218 8.236 10.426 23.105 JAG 11.323 21.759 28.121 8.93 22.849 40.672 CNN 0.029 0.054 0.013 0.039 0.046 0.072 JAG 0.080 0.070 0.036 0.026 0.040 0.030 Table for values of parameter “b” Men 18-35 Men 36-55 Men >55 Women 18-35 Women 36-55 Women >55 Friends 0.029 0.084 0.071 0.035 0.089 0.010 MNF 0.055 0.050 0.068 0.063 0.057 0.033 Malcolm in Middle 0.093 0.085 0.077 0.069 0.061 0.078 Sports Center 0.071 0.094 0.027 0.074 0.055 0.078 TRL Live 0.087 0.018 0.039 0.060 0.014 0.035 Lifetime movie 0.038 0.090 0.051 0.012 0.022 0.050 Cost per ad Cost Friends 160 MNF 100 Malcolm in Middle 80 Sports Center 9 TRL Live 13 Lifetime movie 15 CNN 8 JAG 85 Required exposures Men 18-35 Men 36-55 Men >55 Women 18-35 Women 36-55 Women >55 Required exposures 60 60 28 60 60 28 Model Changing cells: Number of ads purchased Target cell: Minimize total cost Constraints: Total exposures for each demographic group is at least equal to the required exposures Sensitivity analysis How will the optimal cost vary if all the required exposures change by the same percentage? Facility Location models Example 7.7 Warehouse location (WarehouseLocation.xls) Given: Customer location coordinates and shipments per year Customer 1 Customer 2 Customer 3 Customer 4 X-coordinate 5 10 0 12 Y-coordinate 10 5 12 0 Annual shipments 200 150 200 300 Model Changing cells: X and Y coordinate of warehouse location Target cell: Minimize total distance-shipment Constraints: None Sensitivity analysis How will the optimal location change if the annual shipment from a particular customer changes? Modeling issues The distance formula used is for straight line distances. If rectangular distance is more appropriate, use the formula |a-c| + |b-d| for distance, where (a,b) and (c,d) are coordinates. In this example we assumed a warehouses could be built anywhere. But often there may only be a select set of locations where one could be built. Then, we need to incorporate (0,1) programming to this model. Rating teams Example 7.8 Rating NFL teams (NFL2001Ratings.xls) Given: Results of all the games from 2001 season and nominal average rating Model Changing cells: Ratings for each team in the league, home team advantage Target cell: Minimize sum of squared error between actual and predicted point spread Constraints: Average rating is equal to nominal average Terminologies Actual point spread = Home team score – Visiting team score Predicted point spread = Home team rating – visiting team rating + Home team advantage Where predicted home and visiting team ratings must be determined by using VLOOKUP functions Modeling issues Giving larger weight to more recent games may improve accuracy of prediction. This may be achieved by multiplying squared error by k, where k = number of weeks the game is old, and = a factor between 0 and 1.