Asktell/variables objectives

[jlnav]

Introducing specifying variables and objectives to ask/tell generators, This basically supersedes specifying gen_specs.in or gen_specs.out from the user's perspective.

For many of our gens, specify a mapping of variables and objectives to "x" and "f":

variables_mapping = {"x": ["core", "edge"], "f": ["energy"]}

[codecov]

Codecov Report

Attention: Patch coverage is 83.72093% with 14 lines in your changes missing coverage. Please review.

Project coverage is 78.60%. Comparing base (ab09b9f) to head (d66dafb).
Report is 173 commits behind head on experimental/jlnav_plus_shuds_asktell.

Files with missing lines	Patch %	Lines
libensemble/generators.py	76.66%	2 Missing and 5 partials ⚠️
libensemble/utils/misc.py	86.20%	3 Missing and 1 partial ⚠️
libensemble/gen_classes/aposmm.py	66.66%	1 Missing and 1 partial ⚠️
libensemble/utils/runners.py	50.00%	1 Missing ⚠️

Additional details and impacted files

@@                            Coverage Diff                            @@
##           experimental/jlnav_plus_shuds_asktell    #1448      +/-   ##
=========================================================================
- Coverage                                  78.90%   78.60%   -0.30%     
=========================================================================
  Files                                         81       80       -1     
  Lines                                       8077     8054      -23     
  Branches                                    1200     1205       +5     
=========================================================================
- Hits                                        6373     6331      -42     
- Misses                                      1499     1510      +11     
- Partials                                     205      213       +8     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[jlnav]

Dramatically simplified by the user providing a mapping:

    variables = {"core": [-3, 3], "edge": [-2, 2]}
    objectives = {"energy": "EXPLORE"}
    mapping = {"x": ["core", "edge"]}

    gen = UniformSample(variables, objectives, mapping)

and:

    variables = {"core": [-3, 3], "edge": [-2, 2]}
    objectives = {"energy": "MINIMIZE"}
    variables_mapping = {"x": ["core", "edge"], "f": ["energy"]}

    my_APOSMM = APOSMM(
        variables=variables, objectives=objectives, gen_specs=gen_specs, variables_mapping=variables_mapping
    )

Output dictionaries contain the variables, the objective value can be the key in objectives, and the generator uses the provided mapping to repack variables/objectives into "x"/"f".

This is much simpler than my previous attempt at figuring this out dynamically!

[shuds13]

Have you tested if these are in kargs?

[shuds13]

If I forget to comment out my lb/ub line I get:
...persistent_aposmm.py", line 326, in update_local_H_after_receiving
local_H[name][Work["libE_info"]["H_rows"]] = calc_in[name]
ValueError: shape mismatch: value array of shape (4,6) could not be broadcast to indexing result of shape (4,3)

[shuds13]

That was with our branches merged.

[shuds13]

With your branch merged into mine the APOSMM run completes, but I dont have any local optimisation runs happening.

I ran with:

vars_std = {"x0": [mcr, 1.0-mcr], "x1": [-20.0, 20.0], "x2":[1.0, 20.0],
            "x_on_cube0": [0, 1.0], "x_on_cube1": [0, 1.0], "x_on_cube2":[0, 1.0],
            }

aposmm = APOSMM(
    variables=vars_std,
    objectives={"f": "MINIMIZE"},  # SH is "MINIMIZE" being used in APOSMM?
    persis_info=persis_info,
    initial_sample_size=initial_sample,
    #initial_sample_size=100,
    localopt_method="LN_BOBYQA",
    rk_const=0.5 * ((gamma(1 + (n / 2)) * 5) ** (1 / n)) / sqrt(pi),
    run_max_eval= 100 * (n + 1),
    max_active_runs=12,
    dist_to_bound_multiple=0.5,
    #lb=np.array([param.lower_bound for param in varying_params]),
    #ub=np.array([param.upper_bound for param in varying_params]),
    hfun=get_x,
    components=1,
    ftol_rel=1e-1,
)

[shuds13]

Also tried with this, though I think this is the default right.

    mapping = {"x": ["x0", "x1", "x2", "x_on_cube0", "x_on_cube1", "x_on_cube2"]},