wut16 jarosław arabas optimization as a learning...

27
WUT16 Jarosław Arabas Optimization as a learning process

Upload: others

Post on 15-Jun-2020

3 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: WUT16 Jarosław Arabas Optimization as a learning processstaff.elka.pw.edu.pl/~jarabas/ATHENS/wyklad34.pdf · Learning the optimum position Parabola model of the objective function

WUT16Jarosław Arabas

Optimization as a learning process

Page 2: WUT16 Jarosław Arabas Optimization as a learning processstaff.elka.pw.edu.pl/~jarabas/ATHENS/wyklad34.pdf · Learning the optimum position Parabola model of the objective function

Optimization by sampling in R1

?

Page 3: WUT16 Jarosław Arabas Optimization as a learning processstaff.elka.pw.edu.pl/~jarabas/ATHENS/wyklad34.pdf · Learning the optimum position Parabola model of the objective function

Optimization by sampling in R1

Page 4: WUT16 Jarosław Arabas Optimization as a learning processstaff.elka.pw.edu.pl/~jarabas/ATHENS/wyklad34.pdf · Learning the optimum position Parabola model of the objective function

Optimization by sampling in R1

Page 5: WUT16 Jarosław Arabas Optimization as a learning processstaff.elka.pw.edu.pl/~jarabas/ATHENS/wyklad34.pdf · Learning the optimum position Parabola model of the objective function

Learning where to sample

● Using a series of previously sampled points it is possible to predict ● a possible position of the optimum● the most promising area to sample

Page 6: WUT16 Jarosław Arabas Optimization as a learning processstaff.elka.pw.edu.pl/~jarabas/ATHENS/wyklad34.pdf · Learning the optimum position Parabola model of the objective function

Learning the optimum position

Linear modelof the objective function

Page 7: WUT16 Jarosław Arabas Optimization as a learning processstaff.elka.pw.edu.pl/~jarabas/ATHENS/wyklad34.pdf · Learning the optimum position Parabola model of the objective function

Learning the optimum position

Linear modelof the objective function

Expected position of the maximum

Page 8: WUT16 Jarosław Arabas Optimization as a learning processstaff.elka.pw.edu.pl/~jarabas/ATHENS/wyklad34.pdf · Learning the optimum position Parabola model of the objective function

Learning the optimum position

Page 9: WUT16 Jarosław Arabas Optimization as a learning processstaff.elka.pw.edu.pl/~jarabas/ATHENS/wyklad34.pdf · Learning the optimum position Parabola model of the objective function

Learning the optimum position

Parabola modelof the objective function

Page 10: WUT16 Jarosław Arabas Optimization as a learning processstaff.elka.pw.edu.pl/~jarabas/ATHENS/wyklad34.pdf · Learning the optimum position Parabola model of the objective function

Learning the optimum position

Parabola modelof the objective function

Page 11: WUT16 Jarosław Arabas Optimization as a learning processstaff.elka.pw.edu.pl/~jarabas/ATHENS/wyklad34.pdf · Learning the optimum position Parabola model of the objective function

Learning the optimum position

Parabola modelof the objective function

Page 12: WUT16 Jarosław Arabas Optimization as a learning processstaff.elka.pw.edu.pl/~jarabas/ATHENS/wyklad34.pdf · Learning the optimum position Parabola model of the objective function

Learning where to sample

gradient - direction of expected improvement

Page 13: WUT16 Jarosław Arabas Optimization as a learning processstaff.elka.pw.edu.pl/~jarabas/ATHENS/wyklad34.pdf · Learning the optimum position Parabola model of the objective function

Recall steepest ascent method

algorithm steepest ascentinitialize x(0)t ←0while ! stop

d (t)←∇ q(x(t))x (t+1)← x (t )+α d (t )

Page 14: WUT16 Jarosław Arabas Optimization as a learning processstaff.elka.pw.edu.pl/~jarabas/ATHENS/wyklad34.pdf · Learning the optimum position Parabola model of the objective function

Learning where to sample

another direction of expected improvement

Page 15: WUT16 Jarosław Arabas Optimization as a learning processstaff.elka.pw.edu.pl/~jarabas/ATHENS/wyklad34.pdf · Learning the optimum position Parabola model of the objective function

Recall (quasi)Newton methods

algorithmquasi−Newtoninitialize x(0)t ←0while ! stop

d (t)←H−1∇ q (x (t ))

x (t+1)← x (t )+α d (t )

Page 16: WUT16 Jarosław Arabas Optimization as a learning processstaff.elka.pw.edu.pl/~jarabas/ATHENS/wyklad34.pdf · Learning the optimum position Parabola model of the objective function

When models are easy to build

● Objective function is linear

possible to predict the optimum position

● Objective function is a parabola

possible to predict the optimum position

● Objective function is convex (concave)

possible to correctly predict the direction of improvement

Page 17: WUT16 Jarosław Arabas Optimization as a learning processstaff.elka.pw.edu.pl/~jarabas/ATHENS/wyklad34.pdf · Learning the optimum position Parabola model of the objective function

When models are easy to build

● It is possible to conclude on global properties of the objective function by looking at its local behavior (ability to properly extrapolate)

● Series of improvements of the best-so-far result will yield the absolute best in the end

● An analogy to the “best practices”, “quality assurance”

● When you climb up a single hill, you are guaranteed to reach the topmost peak

Page 18: WUT16 Jarosław Arabas Optimization as a learning processstaff.elka.pw.edu.pl/~jarabas/ATHENS/wyklad34.pdf · Learning the optimum position Parabola model of the objective function

When models are easy to build

● It is possible to conclude on global properties of the objective function by looking at its local behavior (ability to properly extrapolate)

● Series of improvements of the best-so-far result will yield the absolute best in the end

● An analogy to the “best practices”, “quality assurance”

● When you climb up a single hill, you are guaranteed to reach the topmost peak

Page 19: WUT16 Jarosław Arabas Optimization as a learning processstaff.elka.pw.edu.pl/~jarabas/ATHENS/wyklad34.pdf · Learning the optimum position Parabola model of the objective function
Page 20: WUT16 Jarosław Arabas Optimization as a learning processstaff.elka.pw.edu.pl/~jarabas/ATHENS/wyklad34.pdf · Learning the optimum position Parabola model of the objective function
Page 21: WUT16 Jarosław Arabas Optimization as a learning processstaff.elka.pw.edu.pl/~jarabas/ATHENS/wyklad34.pdf · Learning the optimum position Parabola model of the objective function
Page 22: WUT16 Jarosław Arabas Optimization as a learning processstaff.elka.pw.edu.pl/~jarabas/ATHENS/wyklad34.pdf · Learning the optimum position Parabola model of the objective function
Page 23: WUT16 Jarosław Arabas Optimization as a learning processstaff.elka.pw.edu.pl/~jarabas/ATHENS/wyklad34.pdf · Learning the optimum position Parabola model of the objective function

When local models may be misleading

● When you want to reach the topmost peak, you should decide which to climb prior to getting on the top

● When you recognize that you have chosen a wrong hill then you should go down and try to find a better goal

● But remember that you are walking without a map and in a fog

Page 24: WUT16 Jarosław Arabas Optimization as a learning processstaff.elka.pw.edu.pl/~jarabas/ATHENS/wyklad34.pdf · Learning the optimum position Parabola model of the objective function

Example model probable location of the maximum

Page 25: WUT16 Jarosław Arabas Optimization as a learning processstaff.elka.pw.edu.pl/~jarabas/ATHENS/wyklad34.pdf · Learning the optimum position Parabola model of the objective function

Example model probable location of the maximum

Page 26: WUT16 Jarosław Arabas Optimization as a learning processstaff.elka.pw.edu.pl/~jarabas/ATHENS/wyklad34.pdf · Learning the optimum position Parabola model of the objective function

Example model probable location of the maximum

Page 27: WUT16 Jarosław Arabas Optimization as a learning processstaff.elka.pw.edu.pl/~jarabas/ATHENS/wyklad34.pdf · Learning the optimum position Parabola model of the objective function

Example model probable location of the maximum