How to write testing files for functions implementing ML algorithms

For a lot of the rule-based algorithms, it is relatively straightforward to write testing files, as there is usually a clear expectation of what the functions should return in various cases.
However, for machine learning algorithms, more specifically, algorithms for fitting probabilistic models, it does not seem to be straightforward, at least to me. Even when we have the generating model with known parameters for data simulation, the model fitting solutions to different simulated datasets will be different, and very often there is no clear expectation of what one should get. For example, if one decides to use Cyclic Coordinate Descent (CCD) for the maximal likelihood estimate of a statistical model, it's hard to know what value one should expect, exactly. And for Bayesian models, the posterior draws are not even deterministic.


Could anyone share any insight/experience?

虽然我没写过，但我觉得是否可以根据expected的结果设定一套，并给出可以接受的的答案范围，这样不管是什么model，只要跑出的结果差的太远就过不了。
不管这个model本身的accuracy以及其他的metrics怎样，test case是根据实际情况设计出来的需要达到的一套标准。
不过，这个test case真的需要很有行业经验的人来设计，不然test case覆盖不全面，或者答案不对的话，过与不过都没什么意义，model accuracy再高也没用啊。