Simulations never converge

[Simon Jeng]

Got it, Thanks Nick and Pasha.

[Carlos Gonçalves]

Hello everyone,

Getting back to the issue of generating a guess file. I came across this procedure cited in the work from Antoine Falisse, Maarten Afschrift, Friedl De Groote https://www.biorxiv.org/content/10.1101 ... 292v1.full

The first initial guess was a "hot-start" based on 172 experimental walking data, whereas the second initial guess was a "cold-start" that consisted of a 173 static, bilaterally symmetrical standing posture that translated forward over the simulation duration. 174 Muscle states and controls were set to constant values for both initial guesses

I think it was exactly what Pasha shared before. The IPOPT and CaSADi can deal with guesses that are within the bounds and path constraints, which doesn't mean that they need to be respecting physics. Is that right? To just leave everything standing and constant was indeed shocking, but definitely is worth trying.

If it is, then creating guess files is tedious but not that hard.

What do you think?

Best regards.

[Aaron Fox]

Hi Carlos,

Yes I don't think the initial guess needs to be anything too 'realistic.' I believe if you do nothing to set the initial guess it simply sets all values to the middle of the bounds on the states across the entire problem - which wouldn't really be respecting physics at all.

Aaron

[Pasha van Bijlert]

Hi Carlos,

Yeah they can get away with "bad" initial guesses, although to be honest I'm not able to reproduce their computational performance. They're describing model complexity way beyond what I'm using, while their 3D simulations seem to converge in under an hour. I'm not sure what explains this difference (I initially thought it might be related to tendon dynamics but I think they're including non-rigid tendons in their simulations).

I'm pretty sure they're using implicit formulations for all the dynamics as well, and using the same contact model, etc. Could the difference lie in the polynomial approximation of the MTC lengths?

Cheers,
Pasha

[John Davis]

I suspect the significant computational speedups are due to automatic differentiation, which is reported to be 1.8-17.8x faster than finite differences for these kinds of problems. My understanding is that Moco currently gets its derivatives from finite differences. The Moco preprint mentions that overhauling OpenSim (and presumably Simbody) to support autodiff would be a pretty substantial undertaking.

[Carlos Gonçalves]

Hello Aaron, Pasha and John. Thanks for the feedback.

@Aaron, that is so true, but yet that is so disruptive for me ... If it works, and will, will be a game changer for my simulations.

@Pasha, I have been following Falisse work for a while, and they use CaSADi directly, and also with "algorithm differentiation" as @John mentioned.

The code for this research is at https://github.com/antoinefalisse/predictsim_mtp.

I "met" some of the researchers from LEUVEN HUMAN MOVEMENT BIOMECHANICS RESEARCH GROUP at ISB 2021 and they use it a lot. I intend to get to learn someway in the next months (fingers crossed).

Hope it helps. Best regards.

Carlos

[Ross Miller]

Algorithmic differentiation can indeed make a tremendous difference in speed for these sorts of simulations. AD is not the same thing as providing symbolic expressions for the gradient but it's conceptually similar in that the time required to determine where to step from iteration to iteration is greatly reduced. I never did much in 3-D pre-Moco but I could do 2-D simulations of a full stride of walking with a fairly complicated model (12 DoF, 26 muscles) with a 101-node grid that often converged within 20 minutes, running Matlab in serial on a normal laptop. That was with symbolic gradients. The drawback to this approach is any time you want to change anything topographic about the model, you have to form the gradient again, which is pretty tedious. There are probably better ways to automate it for someone who is actually a skilled programmer but for me it involved a lot of error-prone copy/paste, find/replace, and grep operations since software for symbolic dynamics like Autolev or PyDy generally doesn't output equations in a format that can be immediately inserted into something like Matlab.

Symbolic is not feasible with OpenSim at least as OpenSim currently exists (it never forms the symbolic state equations) and as John mentioned AD is also not feasible without modifying the "guts" of OpenSim like I think Antoine did in his PhD work. I'm actually impressed that Moco is as fast as it is with finite difference gradients, I was not expecting to be doing 3-D simulations of a full stride with a complex model on my laptop in a single-digit number of hours but that's indeed the performance, most of the time.

The larger headache for me with direct collocation is the number of iterations to convergence is not predictable. Sometimes it is 300, sometimes 3000. Symbolic gradients still suffered from this.

Ton told me once about an approach he was using that worked with random initial guesses but I can't remember the details.

Ross

[Carlos Gonçalves]

Thanks for the complementation Ross.

This is one great forum topic. Hope I can contribute on it later.

It is a long journey ahead is mastering these kinds of simulations.

If anyone is also working on creating guess files, I discovered yesterday that the .sto should have the same columns' positions from what is generated by the solver.

In my case it was:

Coordinates' values
Coordinates' speed
Muscles' activation
Muscles' normalized tendon force
Muscles' excitation (/forceset/soleus_r for example)
Muscles' implicitderiv_normalized_tendon_force


Best regards.