Were the models of Dyer.2022, Grazzini 2017, and Platt 2021 included in our black-it?

[linwenqi2024]

Dyer 2022: Black-box Bayesian inference for agent-based models, Platt 2021: Bayesian Estimation of Economic Simulation Models Using Neural Networks, Grazzini 2017 Bayesian estimation of agent-based models, were the latest advances in the field. It would be appreciated if they were included, and promote largely the value of black-it.

[aldoglielmo]

Hi @linwenqi2024, thank you for opening this issue. Black-it should have already most of the search methods and loss functions used within the Platt 2021 paper as well as Grazzini 2017, have a look at the cors samples , particle swarm sampler , halton sampler and gsl loss , likelihood loss.

We do not currently have the Dyer 2022 methods. To incorporate the methods in that paper, the easiest option would probably be to build an interface to the sbi package https://github.com/sbi-dev/sbi. I can help you out if you give it a try!

[linwenqi2024]

Thanks a lot. if possible, I would like to take a try if you could guide me in details and step by step. I'm a new comer in this field, and it may take a long time to get the results.

[aldoglielmo]

Hi @linwenqi2024,
Well, this is certainly a project where you can learn a lot. If I were you, I would first familiarise myself with both Black-it and the SBI package by running a few tutorial examples from both. Once you understand the basics, you also need to learn the inner workings. Black-it has very intuitive and transparent interfaces (essentially the "Calibrator" object calls alternatively a "Sampler" and a "Loss function", as also described here), but I am not an expert of SBI, and you'll need to understand how that works too.

At a very abstract level, it should definitely be possible to take the building blocks of SBI and conform (some of) them to the "Sampler" and "Loss function" interfaces I just mentioned, so that one can instantiate and call them directly within the Black-it framework