Required `sim_id` key in APOSMM

[roussel-ryan]

This Xopt code fails to add data to the APOSMM generator because it requires a sim_id key in the input data. This is not required by the gest-api standard

vocs = VOCS(
    variables={
        "x0": [-2.0,2.0],
        "x1": [-1.0,1.0],
        "x0_on_cube": [0.0,1.0],
        "x1_on_cube": [0.0,1.0],
        },
    objectives={"f": "MINIMIZE"},
)

variables_mapping = {
    "x": ["x0", "x1"],
    "x_on_cube": ["x0_on_cube", "x1_on_cube"],
}

max_active_runs = 1
initial_sample_size = 10
gen = APOSMM(vocs=vocs, max_active_runs=max_active_runs, initial_sample_size=initial_sample_size, variables_mapping=variables_mapping)

X = Xopt(vocs=vocs, evaluator=evaluator, generator=gen)
X.random_evaluate(10)

End of error traceback

File ~\Documents\GitHub\Xopt\xopt\base.py:392, in Xopt.add_data(self, new_data)
    390 # Pass data to generator, continue in case of invalid data when strict=False
    391 try:
--> [392]  self.generator.ingest(new_data.to_dict(orient="records"))
    393 except DataError as exc:
    394     if self.strict:

File ~\Documents\GitHub\libensemble\libensemble\generators.py:181, in PersistentGenInterfacer.ingest(self, results, tag)
    179 def ingest(self, results: List[dict], tag: int = EVAL_GEN_TAG) -> None:
    180     """Send the results of evaluations to the generator."""
--> [181]  self.ingest_numpy(list_dicts_to_np(results, mapping=self.variables_mapping), tag)

File ~\Documents\GitHub\libensemble\libensemble\gen_classes\aposmm.py:214, in APOSMM.ingest_numpy(self, results, tag)
    212 if self._n_buffd_results == 0:
    213     self._ingest_buf = np.zeros(self.gen_specs["user"]["initial_sample_size"], dtype=results.dtype)
--> [214]   self._ingest_buf["sim_id"] = -1
    216 if not self._enough_initial_sample():
    217     self._slot_in_data(np.copy(results))

ValueError: no field of name sim_id

libensemble/libensemble/gen_classes/aposmm.py

Line 214 in 41a4cbc

self._ingest_buf["sim_id"] = -1

[shuds13]

This should be converted from _id. Is xopt receiving _id from APOSMM, and returning it.

This does show a libE error though, in the conversion looks like libE is modifying the incoming dictionary _id to sim_id

[roussel-ryan]

This is before using APOSMM, arbitrary data not from the generator shouldn't need to provide a _id according to the gest-api

[shuds13]

Okay, we should be able to add that internally.

[shuds13]

I assume the intended behavior here is initial_sample_size is given to tell APOSMM how many results are needed before start optimizing, and the random sampled points are given (ingested) as that (or toward that).

However, APOSMM when initialized currently produces sample points automatically, so we need to update APOSMM so that it can be initialized to not produce that sample. But wait until a suggest is called before producing any points.

[jmlarson1]

This regression test shows how an initial History H0 can be communicated to APOSMM. If that history has enough sample points, then no additional sampling is needed and initial_sample_size can be set to zero.

I'm not sure how to best adjust the regression test to appropriately conform to the gest-api.

[shuds13]

H0 comes in when intialize APOSMM, and as you say, the initial sample is considered generated on top of that. I think it might be more consistent to say initial_sample_size is minimal evaluated points required before starting optimization. Then H0 would count towards that, as would unknown ingested points. But we might want to consider renaming. As APOSMM thinks it must generate this many sample points currently.

Sample phase:

So we want a mode (set on initalisation) to say whether APOSMM is in charge of the sample. In this case, it is not. It will ingest points given into its sample. And if suggest is called it will generate that many sample points (if no number supplied, maybe rest of sample???)

Optimization phase:

Once optimizing, APOSMM will always generate N points for N optimization runs, and fill in more with samples. If suggest asks for less, they get cached. But in this phase, we should enforce that no more unknown points are given via ingest - no more points without _id are allowed.

[jlnav]

As far as the state of our suggest/ingest development branch, I believe this was addressed by #1616