Maximum net size / Multi threading

[Christian]

Hello,

what is the maximum network size I can use? Is it only limited by the 2GB border or also by a maximum number of nodes/weights/arcs?

Is there also a multi thread calculation algorithm? I am currently using the Likelihood algorithm with 10k samples each on 4 cores. After that I calculate the averages of the results. I don't know if that is ok. What do you think?

[shooltz[BayesFusion]]

what is the maximum network size I can use? Is it only limited by the 2GB border or also by a maximum number of nodes/weights/arcs?

2GB is the limit for the user-space memory in the Win32 process. You can use SMILE for 64-bit Linux if you need to get more memory. There are no limits other than available memory for number of nodes/parameters/arcs.

Keep in mind that inference requires additional memory - algorithms create temporary data structures to keep the intermediate results.

Is there also a multi thread calculation algorithm?

No, we don't have an algorithm which uses multiple threads.

[shooltz[BayesFusion]]

I am currently using the Likelihood algorithm with 10k samples each on 4 cores. After that I calculate the averages of the results. I don't know if that is ok. What do you think?

Your approach should work as long as you call srand with different seed in each thread/process. However, I think you should use DSL_ALG_BN_EPISSAMPLING algorithm, which is dramatically better, especially in hard cases (with unlikely evidence).