4. Model of fitness functions ?
No assumption ?
Unrealistically pessimistic results
Unreadable lower bounds
Let's do as in:
==> quadratic
convex objective
functions
6. Complexity upper bounds
Each objective function = a sphere
Below just a short overview of algorithms;
==> the real algorithms are a bit more
tricky
For finding the whole Pareto front :
Optimize each objective separately
The PF is the convex hull
For finding a single point :
Optimize any single objective
7. Finding one point of the Pareto Set
d objective functions
In dimension N
One point at distance at most e of the PF
cost=O( (N+1-d) log (1/e) )
Proof : M log(1/e) in monoobjective
optimization where M is the codimension
of the set of optima (Gelly Teytaud, 2006)
8. Finding the whole Pareto Set
d objective functions
In dimension N
One point at distance at most e of the PF for
the Hausdorff metric
cost=O( (Nd) log (1/e) )
Proof : d times the monoobjective case.
19. How to apply this in MOO ?
Covering numbers can be computed also for
Hausdorff distance ==>
Plus a little bit of boring maths
Leads to bounds as expected
Nd log(1/e) for the whole Pareto set (Hausdorff)
(N+1-d) log(1/e) for pointwise convergence (distance to
one point of the Pareto set)
N=dimension, d=nb of objectives
20. Results in multiobjective cases
The proof method is not new
(Fournier & Teytaud, Algorithmica 2010)
Its application to MOO is new :
Tight bounds
But no result in case of use of surrogate models
(as for corresponding results in the
monoobjective case) ; in fact, the problem
becomes unrealistically easy with surrogate
models...
22. Sorry for not being here
==> really impossible
Discussion ==> all email questions welcome
Tight bounds thanks to a realistic model
Combining previous papers
Complexity bounds Relevant model
Maybe an extension : using VC-dimension
This paper did it
(single objective)