preference modeling and optimization€¦ · our model is a suspension system utilizing a...
TRANSCRIPT
Georgia Institute of Technology
Spring 2008
Preference Modeling and Optimization ME6105 Homework 5 Jonathan Bankston, Yavuz Mentes, Alex Ruderman
Objectives
To develop a deeper understanding of decision making in design
To gain experience modeling preferences in a utility function
To explore a design space and model an optimization problem
To solve design problems computationally and gain a deeper understanding of the computational trade offs.
To develop a deeper understanding of how uncertainty plays a role in design
Background
In HW2, you have identified a particular system for which you would like to solve a design problem. You modeled a
physics-based relationship between design alternatives and design objectives in HW3, and explored the influence of
uncertainty on the system performance in HW4.
You are now ready to solve the entire design problem formulated in HW1. In this final assignment, you will use your
Modelica model and combine it with models for uncertainty and preferences to formulate a complete decision problem.
You will do so in two ways: first, deterministically, to establish a base-line solution, and secondly, considering
uncertainty, i.e., solving the same problem by maximizing the expected utility.
Both the deterministic decision problem and the decision problem under uncertainty will be modeled and solved in the
software ModelCenter by Phoenix Integration.
Tasks
You are asked to write a report on your simulation-based design study. I suggest that you address each of the topics below
in a separate section of your report. Each section should describe your solution to that particular part of the design study,
list the important results (if appropriate), and provide an interpretation of the results.
Important Note: I’m not looking for a 200 page novel... Also, do not hesitate to include screenshots -- a picture is worth
a thousand words. I should be able to grade your assignment based solely on the information in your report without
having to open/run any of your models.
Task 1: Revisit the Decision Situation identified in HW1
Based on what you have learned in this course so far and now that you have a better understanding of the complexity
of your system, briefly revisit the goals and scope of your simulation-based design study.
Include an updated influence diagram that specifically lists the design variables (parameterization of the design
alternatives) and the measures of effectiveness that you will be using in this assignment.
Interpretation: Provide an assessment of what you proposed in HW2 and justify any changes.
The problem we have devised for the semester is slightly different from that described in Homework 2. One
change is the removal of the fluid damper. This is because a traditional damper connected in parallel will
reduce the power output of the system, making it less efficient. We also made some changes to our means and
fundamental objectives hierarchy. The new hierarchies are shown below.
Figure 1.1: Fundamental Objective Hierarchy for Automobile Suspension
Maximize Customer
Satisfaction
Maximize energy recovered
Maximize Safety
Maximize Roll Stability
Maximize Reliability
Maximize Ease of Installation
Minimize Overall Size of Geometry
Maximize Ride Comfort
Minimize Cost
Figure 1.2: Means-Objectives Network for Automobile Suspension
Using the relations from the modified objectives networks above, we made modifications to our original
influence diagram, which is shown below.
Minimize Damping Response Time
Minimize Settling Time
Maximize Power Capability of
LMES
Maximize Energy Recovered
Maximize Ride Comfort
Maximize Customer Satisfaction
Maximize Safety
Maximize Roll Stability
Larger Stator Coil
Thicker Wire More Turns of
the Wire
Maximize number of magnets
Minimize Vertical Acceleration
Minimize Cost
Minimize Cost of PGSA
Figure 1.3: Influence Diagram for Suspension Design
Strength of Magnets
Cross-Sectional Area of wire
# of turns of stator coil wire
Speed of Car
Spring Stiffness
PGSA Capability
Force from Road
Spring Constant
Force to PGSA
Road Condition
Force to Spring
Force to Chassis
Energy Generated
Cost
Comfortable Ride
Vertical Acceleration of Chassis
Mass of Car
Our model is a suspension system utilizing a power-generating shock absorber (PGSA) that we introduced in
Homework 2. The system consists of a spring to counter any vertical motion of the car and the PGSA to soften
the recoil of the spring. The PGSA uses electromagnetic induction to dampen the amount of recoil instead of a
traditional fluid-based damper which uses the compression of a fluid. The PGSA converts kinetic energy from
suspension travel into electrical energy that supplements the electrical system of the car.
The original scenario for this energy-based simulation was the simple vertical motion of the suspension system
installed in a car. The simulations included a displacement representing a single and double bump. Our
original model is shown below in Figure 1.4.
Figure 1.4: Original dymola model of car Suspension with PGSA.
After revisiting this model, we realized a couple flaws. The model we created represents the entire mass of the
car acting on one suspension system. Furthermore, there is no gravity acting on car’s mass. In addition after
reading comments from Chris, we decided that a tire should be included to better represent the system and only
the single bump simulation is necessary for our purposes. With these changes, our models become the figures
below. Figure 1.5 is our new model of the suspension system, and Figure 1.6 is the model of the car attached
four suspension systems and four tires excited at the same time by the same size bump.
Figure 1.5: Existing Dymola Model of the Suspension System
Figure 1.6: Dymola model of the entire car system.
Our objectives are to maximize the ride comfort, maximize electrical energy recovered, and minimize the
overall cost of the system. The ride comfort is characterized by the vertical acceleration of the car where lower
acceleration means a more comfortable ride. The recovered energy from the PGSA is the amount of energy
contributed to the electrical system of the car. We measured AC power from the terminals of the PGSA. The
cost is a relative measurement based on researched costs of the components of the PGSA (i.e. the length of the
copper wire, the cross-sectional area of the copper wire, and the magnetic field for the magnets). These three
objectives are obviously dependent upon one another. Therefore, the attributes were adjusted to find the
optimal system.
Task 2: Elicit Preferences -- Your Utility Function
In order to express your preference with respect to a given design alternative, you need to develop a utility function
that captures the tradeoffs that exist for the attributes of the design alternative. As discussed in class, go through the
elicitation process to determine your utility function. This process becomes quite tedious for more than 4 attributes --
limit yourself to 4.
To document the elicitation process, briefly describe in your report the elicitation questions that you asked yourself,
the fitting of the individual utility functions, the computation of the coefficients in the multi-linear utility function.
Briefly describe how the coefficients in the multi-linear utility function are computed from elicitation questions. In
other words, based on your example, explain the computations that take place in the MAUT spreadsheet.
Interpretation: Provide an interpretation of each of the single-attribute utility functions: Are you "risk averse," "risk
seeking" or both? Explain why? Also provide an interpretation of the coefficients in your multi-attribute utility
function.
As described above, the attributes of our system are the vertical acceleration of the car (ride comfort), the
energy in Joules generated by the PGSA, and the cost in dollars of the PGSA. In order to come up with the
individual utility functions for each of these attributes we used the spline spreadsheet provided by Chris. The
elicitation question that we chose to ask ourselves is shown below.
“At what X would you be indifferent to accepting this X value
or taking a 50/50 chance between X(low) and X(high)?”
In the above elicitation question, X(low) represents the changing value of our attribute corresponding to a
utility value of zero and X(high) represents the changing value corresponding to a utility of 1. The values of X
correspond to the utility value of 0.5 for the different X(low) and X(high) values. In order to change the values
of X(low) and X(high), we made the indifferent X value equal to either X(low) for utilities 0.5-1 and X(high)
for utilities 0-0.5.
Figure 2.1: Monotonic Utility Function for the Total Cost of the PGSA
The curve resulting from the elicitation of our preferences for the total cost of PGSA is Figure 2.1. It was
plotted using the provided Excel spline template. This curve is the monotonic utility function for the total cost.
From the figure, it is evident that we are risk seeking throughout the range of PGSA costs. This means that we
are willing to spend more on the PGSA for increases in performance. It is also notable that the PGSA cost
curve is much smoother than the other two elicitations. We assume that this results from more experience
imagining monetary costs than acceleration or energy generation.
The monotonic utility function for vertical acceleration was elicited using the same question above. This
acceleration acts on the car body and is due to excitation of the lower suspension arms and the weight of the
car. This function was plotted using the same Excel template and presented in Figure 2.2. The general trend of
our preferences towards vertical acceleration appears to be risk-neutral. More specifically, we are slightly risk-
averse in the middle of our range but more risk-seeking toward the end points. This makes sense considering
changes in acceleration near the outer edges do not appear as drastic as changes in acceleration in the mid-
range.
Figure 2.3 is a plot of the monotonic utility function for the energy generated by the movement of the PGSA.
We elicited our preferences for energy generation using the question posed earlier. This elicitation follows the
same pattern as our elicited preferences for the vertical acceleration. Again, we felt more susceptible to change
in the midrange, making us more risk-averse, while we were risk-seeking toward the outer edges of the range.
Figure 2.2: Monotonic Utility Function for the Vertical Acceleration (Comfort) of the Car
Figure 2.3: Monotonic Utility Function for the Energy Generated by the PGSA
We then used the MAUT spreadsheet to generate our multi-linear utility function. To determine the minimum
and maximum utility values for each of the attributes we used the Latin-Hypercube Sampling results from
Homework 4. Using these results helped us estimate acceptable ranges of values for these attributes. Since the
values in the LHS were derived from our input variables that we are confident with, these minimum and
maximum values represent a satisfying range that should not restrict any useful data. We want to use a wide
range of values in order to better gauge the possibilities of our system. From the notes on our graded
Homework 4, we realized that the upper bound of our magnetic field strength distribution values was too large.
We performed the LHS again with a more reasonable range of 0.5 to 2 Tesla for the magnetic field. The
results from the LHS and the elicitations were combined to come up with the multi-attribute utility function.
These process and the results are shown below.
Table 2.1: MAUT Spreadsheet Elicitation
Table 2.1 shows the elicitation section of the MAUT spreadsheet (for three variables) where we elicited our
preferences to calculate the coefficients of our multi-linear utility function. The spreadsheet uses reference
points and elicitation points to determine relationships between the utilities of all three attributes. Our first
step was to develop the list of reference points. A reference point is described by the set of values for each
attribute at that point as well as the corresponding utilities for those values. We retrieved the utility and value
combinations from the monotonic utility functions.
Then, we developed the corresponding list of elicitation points. For each question, we defined two of the
attributes by selecting utilities that were sufficiently far from the reference point. We then found values for
those utilities using the monotonic utility functions. The third attribute was defined by answering the elicitation
question:
“Keeping in mind the range of each attribute and the available
PGSA given by the reference point, what value must the
attribute in question be for you to accept this design given
the predefined values of the other attributes?”
As shown in Table 2.1, we began by repeating the same reference point for sets of three elicitation questions.
This way a different attribute was elicited for each reference point using varied predefined values. If the utility residue or inv(ATA) intermediate results were unacceptable, we adjusted these utilities and elicited the attribute
again.
Elicitation
Elicitation Point Reference Point
Utility
Residue
Attribute #1 Attribute #2 Attribute #3 Attribute #1 Attribute #2 Attribute #3
Question Value Utility Value Utility Value Utility Value Utility Value Utility Value Utility
1 1.75 0.57 1300 0.75 29.7 0.2 2 0.5 850 0.5 26.5 0.5 0.01738
2 0.25 0.875 400 0.017 24.5 0.875 2 0.5 850 0.5 26.5 0.5 0.01589
3 3.25 0.125 1475 0.875 29 0.25 2 0.5 850 0.5 26.5 0.5 -0.01107
4 2.5 0.33 1173 0.675 24.5 0.875 3.25 0.125 1475 0.875 26.5 0.5 0.0461
5 2.75 0.25 1200 0.69 25 0.75 3.25 0.125 1475 0.875 26.5 0.5 0.00903
6 2 0.5 355 0 24.5 0.875 3.25 0.125 1475 0.875 26.5 0.5 -0.02367
7 3.125 0.16 1550 1 34 0 1 0.75 500 0.125 31 0.125 0.01844
8 0.25 0.875 1475 0.875 33.5 0.02 1 0.75 500 0.125 31 0.125 -0.02174
9 1.5 0.64 550 0.25 25 0.75 2.75 0.25 1550 1 24 1 -0.01903
10 2 0.5 900 0.53 29 0.25 2.75 0.25 355 0 24 1 -0.00918
11 0.25 0.875 1300 0.75 30 0.18 2.75 0.25 1550 1 24 1 0.02497
On the spreadsheet, elicitation and reference points represent two design alternatives with equal utility.
Formulas in the spreadsheet are used to exploit this quality to develop a system of equations to describe the
constant utility contours of the multi-linear utility function. Excel functions are then used to solve this system,
producing the multi-attribute utility function coefficients shown in Table 2.2.
Table 2.2: MAUT Spreadsheet Utility Function Coefficients
kVA 0.862892
kGE 0.67316
kC 0.575632
kVA,GE -0.95555
kVA,C -0.84297
kGE,C -0.71852
kVA,GE,C 1.40536
Analysis of the coefficients provides information about the interaction among the design variables.
Specifically, the coefficients representing multiple attributes are important. For example, imagine a coefficient
kXY, developed from random attributes X and Y. If kXY is positive, then the attributes X and Y are
complimentary so our preference for one will increase as the other attribute increases. If kXY is negative, then
the attributes are substitutive, which means our preferences for each change in an inverse fashion. Finally, in
the case of kXY = 0, the attributes do not depend on one another and are therefore considered additive.
From Table 2.2, the values for kVA,GE, kVA,C, and kGE,C are all negative, so it is evident that all of our attributes
are substitutes. This means our attributes are alternates to one another. Our preferences allow for the vertical
acceleration to increase when electrical energy output increases or when the PGSA cost decrease.
Additionally, our preferences accept a lower energy generation for reduced cost and vice versa.
For our PGSA, the multi-attribute utility function was given the coefficient labels in Equation 1. The
coefficients calculated from the preference elicitations in the MAUT spreadsheet were substituted, resulting in
Equation 2.
(1)
(2)
Task 3: Explore the Design Space
Within ModelCenter, implement the entire utility function by combining your Modelica model with "script" models or
Excel spreadsheets. The script models can be used to model measures of effectiveness that cannot be derived from
your Modelica models (e.g., cost models or simple algebraic performance models). The scripts also include the
mappings from individual utility functions into a multi-attribute utility function. The mappings from measures of
effectiveness into single-attribute utility functions is best implemented using the spreadsheets provided.
Explore this design space by using the appropriate exploration tools (e.g., a space-filling DOE) -- the goal is to
narrow down the design space that you will consider for optimization, e.g., to avoid getting stuck in local minima.
These parameter studies will also serve as a validation of your behavior models and utility functions (e.g., it is not
uncommon that a bug in your model surfaces when using parameter values that are different from the ones you
started with when building the model).
For instance, you may do a full factorial exploration of your design space; among the alternatives evaluated in this
full factorial design, find the design alternative that gives you the largest (deterministic) utility. This point could
serve as a starting point for further optimization.
It is also important to develop an idea of how many local minima your utility surface has, whether it is smooth, etc. so
that you can decide on the appropriate optimization algorithm in the next step.
Interpretation: Describe briefly the design exploration steps you performed. What did you learn about your design
problem?
We created a ModelCenter model to explore the design space of our system based on our design variables. As
we show in Figure 3.1, the exploration model consists of two different assemblies: a behavioral model and a
utility model. The behavioral model represents the behavior of our suspension system using the Dymola
model. The design objectives were linked to the corresponding excel spreadsheet for the monotonic utility
functions. The spreadsheet accepts values for vertical acceleration, generated energy, or PGSA cost and
outputs the utility based on the spline curve equation. These individual utility outputs were substituted into the
multi-attribute utility function from Task 2 to calculate the total utility of the given set of input design
variables.
Figure 3.1: (a) Exploration Model, (b) Behavioral model, (c) Utility Model
To explore the entire design space and lower the probability of getting stuck in local maxima during
optimization, we ran a space filling full-factorial DOE with 10 levels. This DOE allowed us to discover the
relationships between the main design variables and the total utility of the system. Evaluating a 10 level full-
factorial created enough runs to generate a smooth surface plot which was used to visualize any local maxima
and determine a good starting point for the optimization process. Since we have three design variables, the
surface plots showed the effect of the magnetic field and wire length on the total utility for each value of the
cross-sectional area, resulting in 10 surface plots. To account for the uncertain variables, we kept the car mass
constant during the full-factorial analysis. We ran the simulation five times to explore design space based on
changes in the car mass (one run for each number of passengers in the car, i.e. 1, 2, 3, 4, or 5 passengers). The
spring constant was kept at the expected value for all five simulations because it was determined previously
that this value did not have a large effect on the design objectives.
Through the design exploration described above, there were five different full-factorial simulations and 10
surface plots per simulation resulting in 50 figures. These plots, along with the individual spreadsheets, were
analyzed to find values of the design variables in a global maximum. In all five simulations, a commonality
with the cross-sectional area was the first few values generated the largest utility. The surface plots for these
cross-sectional area values are shown below.
(b)
(a)
(c)
Figure 3.2: Utility Surface Plot with CarMass=1504 kg and Cross-Sectonal Area=1.0 x 10-6 m2
Figure 3.3: Utility Surface Plot with CarMass=1504 kg and Cross-Sectonal Area=1.5 x 10-6 m2
Figure 3.4: Utility Surface Plot with CarMass=1567.45 kg and Cross-Sectonal Area=5.0 x 10-7 m2
Figure 3.5: Utility Surface Plot with CarMass=1567.45 kg and Cross-Sectonal Area=1.0 x 10-6 m2
Figure 3.6: Utility Surface Plot with CarMass=1567.45 kg and Cross-Sectonal Area=1.5 x 10-6 m2
Figure 3.6: Utility Surface Plot with CarMass=1630.9 kg and Cross-Sectonal Area=5.0 x 10-7 m2
Figure 3.7: Utility Surface Plot with CarMass=1630.9 kg and Cross-Sectonal Area=1.0 x 10-6 m2
Figure 3.8: Utility Surface Plot with CarMass=1630.9 kg and Cross-Sectonal Area=1.5 x 10-6 m2
Figure 3.9: Utility Surface Plot with CarMass=1694.35 kg and Cross-Sectonal Area=5.0 x 10-7 m2
Figure 3.10: Utility Surface Plot with CarMass=1694.35 kg and Cross-Sectonal Area=1.0 x 10-6 m2
Figure 3.11: Utility Surface Plot with CarMass=1694.35 kg and Cross-Sectonal Area=1.5 x 10-6 m2
Figure 3.12: Utility Surface Plot with CarMass=1758 kg and Cross-Sectonal Area=5.0 x 10-7 m2
Figure 3.13: Utility Surface Plot with CarMass=1758 kg and Cross-Sectonal Area=1.0 x 10-6 m2
Figure 3.14: Utility Surface Plot with CarMass=1758 kg and Cross-Sectonal Area=1.5 x 10-6 m2
We learned a lot about our design problem from the design exploration. By separating the analyses by cross-
sectional area, we realized that the utility does not change much for higher values. The utility was greatest for
a smaller range of cross-sections allowing the exploration to become dependent on two design variables rather
than three. We also discovered that there is a significant change in the utility when the mass of the car is
altered due to the number of passengers. However, the largest change was between 1 and 2 passengers and the
variation in utility between 2, 3, 4, and 5 passengers is marginal.
In addition to the trend in utility based on cross-sectional area, we noticed that the global maximum on these
plots was occurring at the maximum bounds of our Magnetic field and Length of Wire design variables. This
meant that our maximum may be too conservative. We were confident with our upper and lower bounds for
the length of wire; however, we were not as confident of our upper bound for the magnetic field. Due to this
we chose to run another full-factorial with a much larger number of levels overnight due to the relatively short
length of time it took to run the 10 level simulation. We chose to increase the upper bound of the magnetic
field to 5 Tesla and run the full-factorial with 40 levels to see if we found a drop in expected utility at the upper
bounds. The increase in levels resulted in 64000 runs taking approximately 12 hours to run. From the
previous exploration, we chose to keep the car mass and spring constant at a their expected values of 1630.9 kg
and 87564 N/m to limit the computation time. The full-factorial generated 40 surface plots (one for each value
of cross-sectional area) and the most promising results are shown below in Figure 3.15-3.17.
(a)
(b)
Figure 3.15: Two view of the surface plot for cross-sectional area=5.0 x 10-7
m2
(a)
(b)
Figure 3.16: Two views of the surface plot for cross-sectional area=6.15 x 10-7
m2
(a)
(b)
Figure 3.17: Two views of the surface plot for cross-sectional area=7.31 x 10-7
m2
We noticed from these plots that there is a drop in total utility resulting in a global maximum (highlighted with
a blue ring) due to our increase in the upper bound of the magnetic field. From this more detailed plot, we
found that the utility increases along the “shelf” up to our upper bound of the wire length. Due to the lack of
time left in the semester, we could not perform a more in-depth analysis to find better bounds for the length of
the wire. Therefore, we used the existing results to determine an optimal design for our PGSA.
From the excel spreadsheet generated with this analysis, we discovered there is a global maximum for our data.
This occurs when the cross-sectional area is 7.31 x 10-7
m2, a magnetic field of 1.88 Tesla, and a wire length of
150 m. We used this as a starting point for our optimization to reduce the possibility of getting stuck at a local
maximum. To attempt to find the global maximum without re-exploring the entire design space, we increased
the bounds on the optimization for the wire length to 500 m.
We were surprised to learn from the exploration that our upper bounds for our design variables were too
conservative. We were confident throughout the semester from our knowledge of the system and the
information gathered from the internet that our bounds were reasonable. After performing an in-depth
exploration of our design space, we felt that an optimum design was possible. The surface plots have clear
local maxima and converge to a global maximum. We were pleased to find from these results that of our
model and utility analysis are accurate. We think that for future work it may be interesting to gain preference
input from more than three people and involve people who have more of experience with this type of
equipment to view the accuracy of our design for an industrial application.
After the design space was explored, we found the optimization algorithm that best fits our system. To do so
we answered the list of questions shown below as presented by Chris in class.
Is the design space continuous or discrete?
To determine if the design space was continuous or discrete, a close look at the design variables and how they
relate to the purchase of the PGSA was needed. The values for the cross sectional area are discrete because the
cross-sectional area of a piece of wire is based on gauge which has set values. The length of the wire, on the
other hand, is continuous since any length of wire can be purchased. The magnetic field based on the method
of purchasing a magnet would be considered discrete because magnets tend to only be sold in discrete tesla
values. Even though two of the three design variables can be considered discrete, we considered the design
space to be continuous to simplify the optimization. We believe this is reasonably accurate because the
magnetic field and cross-sectional area can be rounded to discrete values with marginal effect on the utility.
How many design variables are there?
For the PGSA, we have three design variables: magnetic field, length of wire, and cross-sectional area. In
Homework 4, we determined that there were 2 uncertain design variables that may affect the results: the mass
of the car and the spring constant for the suspension spring.
Is the object function linear or nonlinear?
The objective function we are optimized is nonlinear.
Does the objective function (i.e. expected utility) have local maxima? If so, how many local maxima?
We found from the surface plots that there are local maxima within the design space. From the many surface
plots we generated, it seems there are 3 local maxima.
Is the objective function smooth? Do we have analytical expressions for the derivatives?
The objective function that will be optimized is smooth and the derivative should be relatively easy to
determine.
Are there constraints? If so, are the constraints smooth and convex?
There are no constraints on our model.
From the evaluation of our system design space described above, a good optimization algorithm for our system
was either the Pattern Search or Direct Search algorithm or the Quasi-Newton Methods. These two algorithms
work well for unconstrained objective functions and are meant for local optimization. We believed that the
results would be accurate for these local optimizers as long as we specified starting points close to the global
maxima. Difficulty could have arisen using a Pattern Search or Direct Search algorithm because the local
maxima in our system were not “small” relative to the global maxima.
Task 4: Solve the Design Problem Deterministically
Apply one of the ModelCenter optimization tools to determine the design alternative for which the value is maximized.
Interpretation: Describe, interpret and critically evaluate your result: does the "optimum" seem reasonable? If not,
revisit your models and utility functions.
We used a ModelCenter optimizer to determine which design alternative maximizes the value of our system.
Our ModelCenter license offers two optimization tools: DOT Optimization and Darwin Genetic Optimization.
We chose to use the DOT Optimization because it uses a quasi-Newton method and works well for smooth
constrained and unconstrained optimization problem. Starting with the initial design variable values described
in the previous task, we ran the optimizer for 8 iterations, requiring 61 function calls. Due to our choice to
drastically increase the upper bound of the wire length, we expected a global maximum to be found with at a
higher length than 150 meters. We did not expect much change in the magnetic field or the cross-sectional
area due to the convergence we noticed in the surface plots to optimal values. Our expectations were
confirmed when we used the optimizer to evaluate the design space and showed a maximum utility of 0.8475
to occur for design variable values of 1.853 Tesla for magnetic field, 160.76 meters for wire length, and 6.898
× 10-7
m2
for cross-sectional area. The optimum design is reasonable because it closely reflects the expected
global maximum that was determined from the surface plots in the design exploration. The increase of wire
length above 150 meters verified our belief that our previous upper bound was too conservative. As mentioned
previously future work including a design exploration with the new upper bound for the wire length would
provide more accurate verification of the global maximum
Task 5: Solve the Design Problem under Uncertainty
Create a ModelCenter model that solves the design problem under uncertainty. This involves a nested loop in which
you optimize the results from a probabilistic analysis -- the computation of the expected utility.
Note: this can be very computationally expensive. A few hints to help you reduce the computational complexity:
Use the LatinHypercubeSampling driver rather than a Monte Carlo driver. The LHS driver converges
faster and has the option not to reinitialize the sample between runs. Although a fixed sample will introduce
a bias, it will make it much easier for the optimizer to converge (the objective function in now deterministic)
Solve the optimization multiple times with an increasing number of samples. For instance, start with only 5
samples in your LHS. Once you have reached an optimum, increase the number of samples and restart the
optimization with the previous solution as a starting point. Repeat until the optimum no longer changes
significantly.
Interpretation: Describe, interpret and critically evaluate your results: How is this solution different from the
deterministic solution determined in Task 4? Explain why there is a difference.
We included uncertainty in our ModelCenter model with a Latin-Hypercube Sampling driver. The optimizer
was used to generate values for our design variables while the LHS driver was used to generate a random
sample of our uncertain variables, car mass and the four spring constants. The output of the LHS is the total
utility determined from the random sample of uncertain variables. The utility value from this nested loop is
the expected utility. The optimizer was used to maximize the average value of the generated distribution of
expected utility. The modified suspension model is shown in the figure below.
Figure 5.1: Optimization ModelCenter Model of Suspension System with uncertainty
The design variables for the DOT Optimization tool were the same as in Task 4 (magnetic field, wire length,
and cross-sectional area). The bounds for these design variables were the new bounds described in Task 3.
The bounds for the uncertain variables were the same as described in the previous homework assignments.
The car mass was varied from 1504 kg to 1758 kg based on the possibility of multiple passengers and the
spring constants were varied between 85813 N/m and 89315 N/m based on expected manufacturing defects.
The hypothesis was that the use of these variables in the model optimization loop would determine the set of
our design variables that gives the highest utility for the system and is the optimal design alternative of our
suspension system when considering the effects of uncertainty.
The process of determining this optimal design alternative required many iterations of our design problem
uncertainty model. The optimizer was first run using the optimal result from Task 4 as the starting values for
our design variables with 10 runs for the LHS driver in the nested loop. The optimal design results of the first
run were used as the starting values for the next iteration and the number of LHS runs was increased. This
process was repeated until the variation of the optimal design variables was minimal. The spreadsheet below
is a list of the number runs in the LHS and the corresponding values for each design variables and utility of the design alternative.
Table 5.1: Initial optimization run under uncertainty
#LHS Runs
Magnetic Field (Tesla)
(Start=1.853)
Wire Length (m)
(start=160.756)
Cross-Sectional Area (m
2)
(Start=6.898e-7) # Iterations # Function Calls Utility
10 1.047664 256.2986 1.27700E-06 12 78 0.868
20 0.540463 271.018 1.17105E-06 7 45 0.849
30 0.563595 270.7991 1.18125E-06 4 22 0.8498
40 0.55795 269.9709 1.17992E-06 3 15 0.848715
50 0.556836 268.6767 1.17625E-06 3 15 0.849019
80 0.555835 268.4865 1.17602E-06 3 10 0.849287
As seen in Table 5.1, our use of the optimization process helped our design variables to converge to an optimal
design alternative. The most valuable combination of design variables is a magnetic field of 0.556 Tesla, a
wire length of 268.49 meters, and a cross-sectional area of 1.176 x 10-6
m2. The corresponding utility value for
the nondeterministic case is 0.8493 which is slightly higher than that of our optimal design alternative in the
deterministic case from Task 4.
The values determined for the optimal design of our Power-Generating Shock Absorber when accounting for
uncertainty are well within our bounds. The value for the magnetic field came out to be lower than our
original upper bound of 2 Tesla confirming that our original thoughts were actually correct. The cross-
sectional area value is approximately in the middle of our bounds confirming our value predictions. The wire
length value is within our predicted bounds proving that the values were reasonable. Therefore it is acceptable
to believe that this is a global maximum. However, the design exploration could not be completed to its full
extent so the values actually lie outside of the explored design space due to our change in the upper bound of
the magnetic field. We are confident that the values are preferable due to the expected location of the global
max in the design space exploration and to the high utility value. To better determine the accuracy of this
value would require a larger design exploration to determine that there is not another maximum with the new
bounds.
The optimal design alternative for this nondeterministic case was significantly different than that of the
deterministic case. The large change was from the introduction of the uncertain variables and more
particularly the mass of the car. We knew from both our previous knowledge of electromagnetism and our
sensitivity analysis performed in Homework 4 the car mass had a profound effect on the results of our
objectives and hence a huge influence on our utility. For this reason the large change was not unexpected.
More specifically, the importance of the car mass makes this nondeterministic optimum much more accurate
than that of the deterministic because the car mass based on the number of passengers was not a variable we
could control. Therefore our design needs to be robust for any variation in the number of passengers in the car
to allow for the most energy to be generated and the most comfortable ride created for any situation possible.
Based on the overall design process undertaken, the optimal values were acceptable for the assumptions that
were made. We believe that a more accurate evaluation of the system could be made by expanding the number
of controllable design variables and taking into account other uncertain variables that were not considered.
Due to the lack of high level knowledge and attention required in this course and the short period of time, the
beginning of the problem selection and analysis began a rode of a number of large assumptions that limited the
possibilities of a truly accurate design. Even though the design space was limited from the beginning, the
overall process was followed and internalized by all of us relating in a comparable increase in knowledge of
the work necessary to solve a design problem under uncertainty using the modeling and simulation tools from
this course.
Task 6: Perform Sensitivity Analysis
Finally, perform a sensitivity analysis to see whether your solution changes significantly when you consider epistemic
uncertainty. Select one or two parameters for which you lack confidence in your uncertainty assessment and which at
the same time you expect to influence the overall utility significantly. Solve the design problem again with the
extreme upper and lower bounds for your epistemic uncertainty -- one factor at a time.
Interpretation: Interpret and critically evaluate your results. Is your design decision robust with respect to your
epistemic uncertainty?
To begin the sensitivity analysis, we chose to perform a central composite analysis to determine which
variables have the largest effect on the total utility of the design.
Figure 6.1: Central Composite results for sensitivity analysis
We deduced from the above results that all three design variables and the car mass had a similar level of
significant to the utility. We were confident that our uncertainty analysis of the cross-sectional area was
accurate. We were also confident of the accuracy of our chosen range of values for the car mass, even though
these values were strictly application based because of the large range of vehicle types in the world. We were
not fully confident of our bounds for wire length and magnetic field. For this reason we chose to evaluate the
two variables separately for epistemic uncertainty constraints.
10095908580757065605550454035302520151050
WireLength
MagneticField
CarMass
XSectionalArea
SpringConstant2
SpringConstant
SpringConstant3
SpringConstant1
29%
28%
28%
14%
0%
0%
0%
0%
To do this sensitivity analysis, the optimization procedure described in Task 5 was used while keeping one of
our two epistemic uncertainty variables constant. The design problem was solved once for both the extreme
upper and lower bound of each uncertain variable. This resulted in four spreadsheets that we used to represent
the convergence of the design variables towards an optimum for the magnetic field bounds of 0.5 and 5 Tesla
and the wire length bounds of 20 and 500 meters.
Table 6.1: Sensitivity Analysis for epistemic uncertainty of Magnetic Field
Magnetic Field=0.5 #LHS Runs
Magnetic Field (Tesla)
(Start=0.5)
Wire Length (m) (start=160.756)
Cross-Sectional Area (m
2)
(Start=6.898e-7) # Iterations
# Function Calls
Utility
10 0.5 274.8975 1.25300E-06 5 27 0.8466
20 0.5 281 1.28999E-06 4 18 0.8478
30 0.5 283.391 1.30160E-06 3 13 0.8464
40 0.5 281.295 1.28950E-06 3 12 0.8458
50 0.5 282.456 1.29210E-06 3 12 0.8461
80 0.5 282.456 1.29210E-06 3 9 0.8461
Magnetic Field=5.0 #LHS Runs
Magnetic Field (Tesla)
(Start=5)
Wire Length (m) (start=160.756)
Cross-Sectional Area (m
2)
(Start=6.898e-7) # Iterations
# Function Calls
Utility
10 5 500 5.000000E-06 4 16 0.2155
20 5 500 5.000000E-06 2 3 0.253
30 5 500 5.000000E-06 2 3 0.2329
40 5 500 5.000000E-06 2 3 0.2296
50 5 500 5.000000E-06 2 3 0.2344
80 5 500 5.000000E-06 2 3 0.2328
Table 6.2: Sensitivity Analysis for epistemic uncertainty of Wire Length
Wire Length=20 #LHS Runs
Magnetic Field (Tesla)
(Start=1.853)
Wire Length (m) (start=20)
Cross-Sectional Area (m
2)
(Start=6.898e-7) # Iterations
# Function Calls
Utility
10 2.5856 20 9.24519E-07 4 20 0.784
20 2.5436 20 9.55359E-07 3 13 0.78436
30 2.4815 20 9.66789E-07 3 13 0.78419
40 2.4815 20 9.66789E-07 3 9 0.78449
50 2.4815 20 9.66789E-07 3 9 0.78449
80 2.4815 20 9.66789E-07 3 9 0.78449
Wire Length=500 #LHS Runs
Magnetic Field (Tesla)
(Start=1.853)
Wire Length (m) (start=500)
Cross-Sectional Area (m
2)
(Start=6.898e-7) # Iterations
# Function Calls
Utility
10 1.24735 500 5.000000E-07 5 24 1.18099
20 1.23832 500 5.000000E-07 3 12 1.18486
30 1.24423 500 5.000000E-07 3 13 1.18914
40 1.237488 500 5.000000E-07 3 13 1.1894
50 1.24073 500 5.000000E-07 3 9 1.188
80 1.23045 500 5.000000E-07 3 12 1.1889
By fixing these two variables at their extreme bounds we were able to see the robustness in our design based
on our lack of knowledge of the uncertainty of these variables. As seen in Table 6.1, there was no real change
in the design variables at the lower bound of the magnetic field due to the closeness of the optimal value to this
bound. At the upper bound, the other two variables also converged to their upper bound. The optimal utility is
also much lower. This drastic change in optimal values was due to the high increase in cost for a higher power
magnet causing a requirement of a high amount of generated energy and a lower vertical acceleration. We
concluded from these results that the overall design of the PGSA was not robust to major changes in the
magnetic field. This conclusion was expected and is always a concern when designing systems that utilize
magnets. The magnet was usually the driving force for any fluctuation in cost.
The epistemic uncertainty evaluation of the wire length is shown in Figure 6.2. When the wire length was set
to the lower bound, the optimal cross-sectional area did not change much, but there was a large jump in
magnetic field due to the large drop in wire length. From our knowledge of electromagnetism, we suspect that
there is a direct relationship between both wire length and magnetic field and the amount of energy generated.
Therefore we expect that a drastic drop in the wire length requires a much more powerful magnet to have a
comparable utility for the design. At the upper bound, the magnetic field increased to improve the utility of the
system. The cross-sectional area dropped to its lower bound which indicated an attempt to lower cost due to
the increased cost of the longer wire. The utility of this optimum was also much higher than the preferred
design alternative derived in Task 5.
From this sensitivity analysis, it was clear that the preferred alternative was highly sensitive to the epistemic
uncertainty due to both the dramatic changes in the design variables and the total utility when the uncertain
variables were changed. We determined through this evaluation that the system is not robust relative to the
epistemic uncertainty. This is the same case in fluid damper suspension system. The design specifications are
specific to the car used. In the future, more and more cars will benefit from such a power-generating source as
our shock-absorber, so a wide range of cars and trucks can use this design. For these reasons we expect that
the design of the PGSA is application based.
Task 7: Lessons learned
In a few paragraphs, identify what you learned in this assignment. Where did you struggle? Where did you go down the
wrong path? How did you revise your approach? If you were to solve another design problem, how would you do things
differently? Did you learn anything new about the trade-offs in the system you were designing?
Jonathan Bankston
In this homework assignment, I was introduced to the amount of work involved in the optimization of
variables in an engineering design. Many people, including myself, like the throw out the term optimization
loosely, as if it was a simple process with clear-cut results. I’ve become very interested in sensitivity analysis
and their possible application in design coordination. Completing this assignment helped me learn more about
performing a sensitivity analysis. I also learned what can and cannot be learned through behavioral modeling. I
want to explore the use of ModelCenter in my future research, particularly in linking ModelCenter with the
informational databases available in electrical CAD software.
Yavuz Mentes
After finishing all the assignments in ME 6105, I am delighted to easily say that I gained at least the basic
knowledge about modeling and simulation which will help me in my Structural Engineering career. My goal
was getting familiar with Mechanical Engineering and systems realization by taking modeling and simulation
course so I ended up taking ME 6105 after I checked its course content. As we started working on assignments
1 and 2, I learned how to use Dymola and started thinking beyond my original civil engineering view.
Actually, the topics we covered after we started working on our group project are going to be useful for my
future career. Especially, uncertainty analysis and sensitivity analysis were the most amazing topics. I learned
how uncertainty of different variables could affect the whole system. In addition to this, in this last homework,
I learned how we can reach an optimal value for different variables that would affect the whole system. The
effect of uncertainty on optimization was an advanced learning experience for me. Overall, I gained a great
amount of value from ME 6105 and I am sure it will pay off in the future...
Alex Ruderman
By performing all of the steps required in the complete design of our system I learned a great deal about utility
properties and the process of optimization. Due to the relative nature of this process to my thesis research, I
gain much value from the lessons learned about the iterative process of design space exploration and
simulation iterations. I realized in Task 3 that the design exploration portion of the process can be very
tedious. I know in our exploration we ran into problems with our bounds and the fact that we had three design
variables. Since a surface plot can only show the relationship between 2 variables and utility, it became very
cumbersome to have to extract plots for all of the variations in the third parameter and interpret them. We
ended up going down the wrong path at first because we did not take into account that the bounds could be too
conservative. When I realized this after talking to Chris, we were forced to redo the design exploration at the
last minute. For me this part became very frustrating because of the lack of time to generate the most accurate
results. I think that by simply increasing the bounds on the wire length without actually performing the
exploration cut a lot of the necessary time down and gave a reasonably accurate result. If I had to do it over
again I think we should have ran smaller full-factorials (maybe level 5 or 10) that would give a small picture
and then increase the bounds and run a higher level analysis later. Overall I feel that I gain much value from
this course project for my research and enjoyed using Dymola and ModelCenter for the design problem
evaluation.
Task 8: Project Web-Page
Add a short paragraph to your project web-site, maybe with a screenshot of your ModelCenter model, illustrating the
outcomes of your simulation-based design study. You should also upload, in a zip folder, all the files you need to run the
ModelCenter model: all the Excel, *.pxt and *.pxc files, and the folder with the final file wrapper and corresponding .mo
file.
Final Project Presentation: Thursday May 1st at 11:30AM.
During the final exam period, we will provide each (group of) student(s) with an opportunity to make a 10 minute
presentation about their project.
The goal here is to learn from each other. Shortly introduce your application domain and then highlight the most
interesting aspects of your assignments/project: Where did you encounter unanticipated obstacles, and how did you
overcome them? Did you learn any neat modeling tricks you want to share with your colleagues? Did you obtain some
interesting results? Can you provide us with some interesting insights based on your models?
So, the focus should not necessarily be on the results of your design study, but on the interesting issues that you
encountered in the process of solving your design problem. This Project Presentation is meant to be a fun learning
experience -- don't hesitate to entertain some also.
Evaluation Criteria
Is the design problem well-defined?
Are the individual steps in defining the design problem performed correctly?
How well are the results interpreted and critically evaluated?
Project presentation: contents and presentation (only small portion of grade)
Organization and clarity of thoughts and report
Submission
1. Hard copy in class (for on-campus students only)
2. T-square: upload your report and all your Modelica (*.mo), Excel (*.xls) and ModelCenter files (*.pcx,
.fileWrapper) to t-square. All your files should be included in one zip file. Name the zip file:
ME6105.HW4.yourfamilyname.zip. (for all students)
Distance Learning Students: There is no need to fax your hard copy in to DLPE. It is easier for me to just download it
from t-square.