Page 1 of 1

Vensim optimization / Powell’s “Hill Climbing” algorithm

Posted: Fri Jan 17, 2020 9:12 pm
by Sarah88
Dear all,

I would like to describe in my dissertation how the vensim optimization option (calibration) works in technical terms (not going into great detail, but providing transparency on this).

In particular, I am interested how the payoff function is calculated, what rule(s) decide that Vensim stops searching for better solution, what happens in the existence of several optimal local optimum and how the size of the marginal changes (in the selected parameters that may vary) are determined? I haven't found a reference where these questions are clearly answered and an explanation is provided.

Many thanks for any answers/explanations or useful references! (there has been a previous discussion on this, but the indicated reference is not available any more i.e. the webpage doesn't open).

Very best,

Sarah

Re: Vensim optimization / Powell’s “Hill Climbing” algorithm

Posted: Mon Jan 20, 2020 5:37 pm
by tomfid
It's fairly close to the Numerical Recipes version:
http://www.nrbook.com/a/bookcpdf.php

Powell's method (and any other hill climber) will find only a single local optimum, unless you do a grid or random variation of the starting point. Then the probability of discovery of additional optima is proportional to size of their basin within the search space.

Simulated annealing does guarantee discovery of a global optimum if your cooling schedule is logarithmic, which usually requires a completely impractical amount of time. However, it can be an interesting experiment, because a stochastic search can also avoid other pathologies, like getting stuck in a narrow twisting valley. https://metasd.com/2018/05/optimization-banana-death/