Added problem hessians and tests #104
Conversation
|
Hm. For some reason your git is messing with the file permissions. I haven't seen that issue before, but have a look here. Maybe that can fix it. |
|
Can I ask where you can see that there is a problem? 1) so I can see the error message and 2) so I can see if it is fixed if I try fix it |
|
If you take look here https://github.com/JuliaOpt/Optim.jl/pull/104/files then after each file name it says 100644 → 100755 which is showing the change in file permissions. |
|
I have no idea why that happened. I just chmod'ed, committed, and pushed. Now it should be fine, right? |
|
Looks much better. |
| @@ -77,70 +87,75 @@ examples["Fletcher-Powell"] = OptimizationProblem("Fletcher-Powell", | |||
| fletcher_powell, | |||
| fletcher_powell_gradient!, | |||
| fletcher_powell_hessian!, | |||
| [0.0, 0.0, 0.0], | |||
| [[0.0, 0.0, 0.0]], # TODO: Fix | |||
| [-1.0, 0.0, 0.0], # Same as in source | |||
There was a problem hiding this comment.
So this works now with Newton's method?
There was a problem hiding this comment.
No, because the Newton tests are only run for the problems with twicedifferentiablefunction == true.
There was a problem hiding this comment.
Optim.nelder_mead finds the optimum, but I'm not sure what would happen with a method based on the gradient or Hessian. The gradient isn't really defined at the optimum I guess, since d atan/dx is 1/(x^2/y + y), and the solution is x=(x,y,z)=(1,0,0), and you would also have a 1/y'ish term in the hessian.
|
Minor nit: we don't put spaces around |
|
I've changed the ^ locally, and they'll be fixed when I push the next time. |
|
So I changed the ^-spacing, and I also changed the initial value from [2.8,1.8] to [2.0,1.5] just to start a further away from [3,2]. To be honest I don't think it makes sense to think too much about what the starting values should be in these test problems, until the problem with starting in concave regions is resolved. Whenever I seem to start at a concave region, I get errors. I wouldn't expect this to happen, though, but I have to read a bit more into how Optim's Newton-algorithm is globalized. |
|
I don't think there's anything special done to globalize Newton's method. |
|
Okay, so in that case I guess it actually should throw an error if I start it in a concave region, right? I mean, the direction is going to be uphill, and without a correction to the hessian, the linesearch isn't going to help. |
|
@johnmyleswhite Do you have further comment here? If not, I think it can be merged when the commits have been squashed. |
|
No more comments. Merge ready now for sure. |
permission change permission change Changed space around ^
|
Squashed and pushed... Let's see what Travis has to say! |
|
@johnmyleswhite Can you push the bottom? I'm not authorized to that. |
Added problem hessians and tests
PR as a follow-up to #101 where I messed up the git history