L14 - Multi-objective (Evolutionary) Algorithms Continued Flashcards
What relation do we use to establish the Pareto front?
Dominance relation
How can we use tournament selection with dominance criterion for multiple objectives?
- Choose points X and Y and check for dominance
- If yes, no poblem
- If no, we choose reference set of points Z1…ZN
- Check how many Z’s X and Y dominate
- Select the X or Y that dominates the most Z’s
In selection of multiple objectives using dominance criterion, what is C? What’s its impact?
- C is the number of Z’s that are selected for reference for X and Y
- Greater C = Larger scope of Z’s and thus more chance of X or Y dominating the other.
What is the issue with tournament based Pareto front selection? What is a solution to this?
- 2 points can be mutually non-dominating, leading to ties.
- Tiebreak resolution
What is niching?
- A solution to tiebreak in dominance base Pareto selection
- Create a selection contained in a niche radius around each solution wen might keep.
- Niche radius is a parameter we choose
- Prefer niches with few solutions
What are the types of Tiebreak solution?
- Solution spread
- Niching
What is Solution Spread?
- When there is a tiebreak during dominance based Pareto selection, choose the point in the least densely populated area?
What is an issue with niching?
- Niche Radius is a net parameter we need to set
- Problem specific parameter
What is Crowding Distance? Why was it created?
- A distance measure that removed the need for the niche radius
- Compute the distance between each solution and its nearest neighbour and great a square around each point on the Pareto front
- Choose the most sparsely populated square
What is the difference between an Elitist and Non-Elitist selection?
Elitist -> Combine best solutions across generations
Non-elitist -> Only have best solutions from most recent generation
What is an issue with Elitist selection?
- Can cause premature convergence in a local optima
What is an issue with non-elitist selection?
- Discards potentially excellent solutions in parent generations.
What are the issues with Many-Objective Optimisation?
- Visualisation issues with a high dimension Pareto front
- Reduction in search capability
- Exponential increase in the number of solutions needed for a Pareto front