MAP queries and learning of deterministic circuits

[mattShorvon]

Hi guys

Hope you're well! I'm looking to use deterministic PCs to run MAP queries on. Are methods of structure learning for deterministic circuits going to be implemented, for example the Strudel method? Correct me if they already are implemented, I may not have looked hard enough. Furthermore, I can see code to run marginal and conditional queries, but is there/will there be code for MAP queries implemented soon? If not, would it be easy to implement if I wanted to give it a go myself?

Kind regards

Matt

[liuanji]

Hi Matt,

I hope you are well as well! No, we currently don't have an implementation for MAP, but we would like to add this feature. What features would you need since you also mentioned algorithms to learn deterministic PCs (e.g., Strudel)?

Best,
Anji

[mattShorvon]

Hi Anji

Thanks for your quick reply! Those are the only two things really, a function for MAP queries and a function for structure learning of deterministic circuits. I mentioned Strudel as I believe that is the current SOTA for learning deterministic circuits. A function for MMAP queries, or the ability to specify variables to marginalise over in the MAP function as an extra argument could be useful too, and it could give an exact solution if the input circuit is q-deterministic. Following on from that, I'm not sure if the Choi et al. pruning method for making circuits that can do MMAP is implemented (https://proceedings.mlr.press/v151/choi22b.html)? Not as important for me though, the main two things are just MAP() and strudel.

Kind regards

Matt

[liuanji]

Hi Matt,

The MAP query is more straightforward and requires less effort to implement on our side. I'll put it on my TODO list and get it done ASAP. Implementing structure learning approaches like Strudel would be harder, but I can investigate. I guess one good option is to use the code based on Juice (https://github.com/UCLA-StarAI/Strudel). You could then choose whether to do MAP inference directly in Juice or load the same circuit in pyjuice once the MAP query is implemented.

Best,
Anji

[mattShorvon]

Thanks Anji! I've had a look at ProbabilisticCircuits.jl, and might not have looked hard enough but couldn't find a Strudel implementation in the latest version. I think its only implemented in v0.3.3 and before, so I've been seeing if I can get an old version up and running. Thanks so much for your help tho and looking forward to seeing the MAP query function!