User report

[mrakitin]

@lyang11973 reported an issue with this optimization problem:

pos_x10 = crl.x1.position
pos_x20 = crl.x2.position
pos_y10 = crl.y1.position
pos_y20 = crl.y2.position
search_range = 0.6

dofs = [
    DOF(device=crl.x1, description="transfocator upstream x", 
        search_domain=(pos_x10-search_range/2, pos_x10+search_range/2)),
    DOF(device=crl.x2, description="transfocator downstream x", 
        search_domain=(pos_x20-search_range/2, pos_x20+search_range/2)),
    DOF(device=crl.y1, description="transfocator upstream y", 
        search_domain=(pos_y10-search_range/2, pos_y10+search_range/2)),
    DOF(device=crl.y2, description="transfocator downstream y", 
        search_domain=(pos_y20-search_range/2, pos_y20+search_range/2))
]

objectives = [
    Objective(name="em1_sum_all_mean_value", description="beam intensity", target="max", trust_domain=(100000, np.inf))
]

dets = [em1]

agent = Agent(dofs=dofs, objectives=objectives, detectors=dets, db=db)

RE(fast_shutter_wrapper(agent.learn("qr", n=16))) 

It completes the first set of data collection, but then fails with an error:

TypeError: where(): argument 'other' (position 2) must be Tensor, not float 

Proposed fix

In agent.py: np.nan needs to be converted to torch.tensor(np.nan).

@jennmald, @yxrmz, please have a look. We need to:

• implement the fix
• add a test for the use case
• update the doc string to explain the enforce_all_objectives_valid parameter