Notes on Value Function Iteration

1

A very nice set of lecture notes by Karen Kopecky on value function iteration, publicly available on her webpage:

Out of couriosity, I checked the methods she suggests to solve VFI numerically and the toolkit has almost all of them!

  1. Brute force (called pure discretization in the VFI toolkit convention)
  2. Howard improvement (it is the default in infinite horizon models, not applicable for finite horizon)
  3. Concavity of value function (missing)
  4. Monotonicity of policy function (called ‘divide and conquer’, set vfoptions.divideandconquer=1)
  5. Linear interpolation (called ‘grid interp layer’, set vfoptions.gridinterplayer=1). The interesting thing is that the toolkit implementation is really close to Karen’s pseudo-code.
1 Like
2

High quality notes :slight_smile:

One difference that is both important and minor: toolkit does these on GPU, rather than CPU, so you have to reimagine all the details of implementation so that things are in parallel wherever possible. So same things being done in terms of headline concepts, but with substantial differences in terms of exact implementation details.

2 Likes