convergence problem when fitting complex model with multiple regression

[XiaoyuZeng]

Hi all.
Many thanks for developping such an incredible toolbox. I used to struggling with the between-subject within-subject interaction in HDDM, but now, HSSM makes it so much easier to examine such interactions!

yet, I had this convergence problem when I fitted a ddm with multiple regression that include complex interaction for all of the four ddm parameters. The trace plots was a mess and suggested that most of the parameters did not converge across 4 chains, and the effective samples and r hat were also far from ideal.

I am aware that I can simplify the model, incraese samples and burn-ins (I noticed hssm does not provide thin argument, correct me if I was wrong about this), and also even consider just fitting group-level parameters.

my question is, is there anything in hssm (like particular sampler?) can help to improve the convergence problem? and maybe also tips to speed up the fitting process (like gpu?)

model specification:

mfull = hssm.HSSM(
model="ddm",
data=df,
include=[
{
"name": "v",
"formula": "v ~ 1 + condition * stage + (1+condition * stage|subj_idx)",
"link": "identity",
},
{
"name": "z",
"formula": "z ~ 1 + condition * stage + (1+condition * stage|subj_idx)",
"link": "identity",
},
{
"name": "a",
"formula": "a ~ 1 + condition * stage + (1+condition * stage|subj_idx)",
"link": "identity",
},
{
"name": "t",
"formula": "t ~ 1 + condition * stage + (1+condition * stage|subj_idx)",
"link": "identity",
},
],
)

sampling setting:

cores=20, # how many cores to use
chains=4, # how many chains to run
draws=10000, # number of draws from the markov chain,
tune=5000, # number of burn-in samples
idata_kwargs=dict(log_likelihood=False), # return log likelihood

example screenshot of trace plot

example convergence diagnosis

Best,
Xiaoyu

[XiaoyuZeng]

hi All,

I happened to find that I can set target_accept argument in the hssm.sample,

for now, I set the target_accept = 0.9. I found that hugely slowed down the whole sampling procedure. I was wondering can i reduce the number of draws and tune as I already restrict the target_accept?

Do you have any advice to optimize the setting of target_accept + draws + tune?

Best,
Xiaoyu