Project Evaluation¶
It is nearly impossible to derive the “best” choice. Therefore, we try to find the “preferred solution”
What is “best”?¶
Extreme (high/low) of all possibilities
Either
- 1 metric of performance
or
- Metrics can be put on single scale
However, both 1 and 2 are not realistic
Value/Preference/Utility Function¶
\(V(x)\) is a means of ranking the relative preference of an individual for a bundle of consequences \(x\)
Diminishing marginal utility curve¶
Exceptions to Diminishing Marginal Utility¶
Very common in real life
- Critical mass: only valuable if you have enough
- Network/Connectivity: more connections \(\implies\) more valuable
- Threshold/Competition: only valuable if
- Minimum reached (absolute graded exams)
- Matches/beats competition (relative grading exams)
Conditions for a Value function¶
Axioms¶
- Completeness/Complete Pre-order: \(V(x)\) is defined \(\forall x_i\)
- Transitivity: \(V(a)>V(b) \ \land \ V(b)>V(c) \implies V(a)>V(c)\)
- General true for individuals
- Not necessarily true for groups; not all group members share the same preferences
- Ellsberg Paradox: Under ambiguity, transitivity does not always hold, as people will want to choose the non-ambiguous option usually
- Allais Paradox:
- Monotonicity/Archimedean Principle
- \(V(x)\) is monotonically-increasing/decreasing
- \(a > b \implies (V(a) > V(b) \quad \forall a, b) \lor (V(a) < V(b) \quad \forall a, b)\)
- This assumption does not hold for all utility functions
- Inflation rate
- Audio volume
- Salt on food
- Problem can be re-formulated as “Salt available on table”
Consequences¶
- Existence of \(V(x)\)
- Only ranking \(x_1, x_2, \dots\) possible. We cannot quantify the distances between \(V(x_1), V(x_2), \dots\)
- Strategic equivalence: Monotonic transformation of \(V(x) \equiv V(x)\); \(V(x_1, x_2) = {x_1}^2 x_2 \equiv 2 \log \vert x_1 \vert + \log \vert x_2 \vert\)
- Values not good basis for absolute value
- Arrow’s Impossibility Theorem/Paradox
- No “fair” voting system, without a dictator, that satisfies everyone’s preferences
- Hence, concept of “best” is not meaningful in design of complex systems
- Therefore, we try to find the “preferred solution”
Outcomes¶
Nature of Evaluation
- Many dimensions & metrics of performance
- Uncertainty about metrics
- “Best” is undefined
- We can screen out dominated solutions
Nature of Choice
- Any person must make tradeoffs
- Group inevitably have to negotiate deal
Concept of Dominance¶
One alternative better than others on all dimensions
Dominated alternatives can be discarded
Feasible region or “Trade Space” is area under & left of the curve
Metrics¶
- Expected Value: Useful, but insufficient, as it cannot describe range of effects
- Worst-case scenario with some notion of probability of loss: People are “risk-averse”; more sensitive to loss
- Best case scenario
- CapEx: Capital Expenditure = Investment
- Some measure of benefit-cost
- Value-Modelling
- VAR
- VAG
Robustness¶
Taguchi method
Robust design is a product whose performance is minimally-sensitive to factors causing variability
Robustness measured by standard deviation of distribution of outcomes
Preferred when we particular result
- Tuning into a signal
- Fitting parts together
However, this is not necessarily value maximizing. We would prefer to
- limit downside
- maximize upside