Rename HMC to LMC about klara.jl HOT 4 CLOSED

If it is strictly equivalent, then why not go for a type alias HMC = LMC ? There would be a single implementation needed.
I admit that the HMC sampling was the only one familiar to me. I am uneasy about calling it LMC as it is an adaptation from a paper that was calling it Hamiltonian Monte-Carlo.

from klara.jl.

Lagrangian Monte Carlo is not a novel term :) For example, look at this paper where it is used. The implementations differ between LMC and HMC, and in fact there can be differences in performance depending on the problem on which the samplers are run. Have a look at this useful link too.

I do not insist that we rename it to LMC, yet at the same time it is sth established in literature and it will have to be used for RMLMC (Riemmanian manifold Lagrangian Monte Carlo). Perhaps we can find another way of dealing with it, avoiding to rename the existing HMC type. Send me the link to the article you read when you have the time (I usually reference Neal's HMC paper) ;)

from klara.jl.

The HMC algorithm I implemented is from the same paper introducing the NUTS algo : http://arxiv.org/abs/1111.4246. They do call it Hamiltonian Monte-Carlo. Is that a case of people using different names for the same thing ? I am a bit confused !

Thanks for your links, I'll check these out.

from klara.jl.

Hi @fredo-dedup, I looked at the paper you sent me. Yes, algorithm 1 in page 3 is HMC, not LMC. This means that the HMC sampler does not need any renaming. I also looked at your code and you implemented HMC indeed. My initial confusion came from the fact that you called the auxiliary variable velocity in the HMCSample type - this is the only change we need to make, simply to rename this field to momentum. I am going to close this issue now.

from klara.jl.