ML Wiki
Machine Learning Wiki - A collection of ML concepts, algorithms, and resources.

Ideal Point Model

Ideal Point

In a Multi-Objective Optimization Problems there are a set of the “best” (Pareto-optimal) solutions. How to decide which one to take?

The ‘‘ideal point’’ (or ‘‘datum point’’) is a solution that is not feasible, but most desired:

  • we take all the extreme solutions
  • and take the values of criteria at which they are best
  • from these criteria we form the ideal point $i$
  • Image

Then we use the weighted distance to compute a point which is closest to the ideal

The distance function is:

  • $\min \left[ \sum_{k=1}^m w_k (z_k(x) - i_k)^p \right]^{1/p}$
  • this is called the $LP$-distance function

Exercises

Exercise 1

Consider the following multi-criteria problem:

  • $\max X_1 + 2 X_2$
  • $\max X_1$

And the following constraints:

  • $X_1 \leqslant 2$
  • $X_1 + X_2 \leqslant 3$
  • $- X_1 + X_2 \leqslant 1 $
  • $X_1 \geqslant 0, X_2 \geqslant 0$

Compute

  • the efficient solution set
  • the datum point
  • find the closest points to the datum point with $L_1$ and $L_2$ distances

First we construct the solution space:

  • Image
  • the efficient solution set is the one that maximizes everything
  • we select the following points: $A(1, 2), B(2, 1), C(2, 0)$
  • values that lay on line $ABC$ form the efficient set of solutions

Now we draw the criteria space

  • we build it from all possible values of the line $ABC$
  • $P(X_1, X_2) \to P(X_1, X_1 + 2 X_2)$
  • $A(1, 2) \to A(1, 1 + 2 \cdot 2) = A(1, 5)$
  • $B(2, 1) \to B(4, 2)$
  • $C(2, 0) \to C(2, 2)$
  • the ideal point is $D$: a virtual point that would be best for both criteria
  • $D(5, 2)$
  • Image

Now we calculate which solutions are closest to $D$:

  • for $L_1$: all values on the line $AB$
  • for $L_2$: the middle between $A$ and $B$

Sources