Skip to content

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. 1 metric of performance

or

  1. 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

image-20240201215653668

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

  1. Completeness/Complete Pre-order: \(V(x)\) is defined \(\forall x_i\)
  2. Transitivity: \(V(a)>V(b) \ \land \ V(b)>V(c) \implies V(a)>V(c)\)
  3. General true for individuals
  4. Not necessarily true for groups; not all group members share the same preferences
    • image-20240201225315043
  5. Ellsberg Paradox: Under ambiguity, transitivity does not always hold, as people will want to choose the non-ambiguous option usually
  6. Allais Paradox:
  7. Monotonicity/Archimedean Principle
  8. \(V(x)\) is monotonically-increasing/decreasing
  9. \(a > b \implies (V(a) > V(b) \quad \forall a, b) \lor (V(a) < V(b) \quad \forall a, b)\)
  10. 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

image-20240201230949309

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

image-20240201233033845

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

image-20240201233230356

Last Updated: 2024-05-12 ; Contributors: AhmedThahir

Comments