Lecture 7 Flashcards
what is the convergence velocity?
the expectation of the distance towards the optimum covered per generation
phi = E[|f(x) - f(xt)| - |f(x) - f(x{t+1})|]
what is the 1/5 success rule?
the optimal success probability should be about 1/5. If, during execution of (1+1)-ES, it is larger than 1/5, increase sigma, if smaller than 1/5, decrease sigma
for the corridor and the sphere model, what are the optimal standard deviations?
corridor: sigma’_opt = sqrt(2pi)
sphere: sigma’_opt = 1.224
what is the corridor model?
corridor represents a range in the solution space where the algorithm focuses its search
- the width influences the exploration behaviour of the algorithm
- as the optimization progresses, the corridor narrows
- wider corridor => slower convergence
what are the maximum convergence velocities for the corridor and the sphere model?
corridor: 1/e
sphere: 0.2025
what are the optimal success probabilities for the corridor and the sphere model?
corridor: 1/(2e)
sphere: 0.27
what is the sphere model?
a simplified optimization landscape with a convex, unimodal objective function that provides an environment for evaluating the performance
what is the formula to calculate the convergence velocity for the sphere model?
phi = E(R^2 - r^2)
R is the radius of the sphere that represents all the possible definitions of the parent, r is the radius of the sphere that represents all possible definitions of the offspring
why would you measure the fixed-target running time?
measure the cost, quantitative comparison
what is the advantage of assessing the fixed-budget objective value?
directly applicable to runs with a very small number of function evaluations
how do you calculate the expected running time of an algorithm?
divide the amount of evaluations by the success rate
what is neuroevolution?
the process of automatically configuring and training artificial neural networks using EAs
part of Neural Architecture Search (NAS)
what are the tree main components of NAS models?
- search space
- search strategy
- performance estimation strategies