batch>1 does not work for local_penalization

[elzurdo]

GPyOpt version 1.2.1 python3 version : 3.6.5

I would like to run suggest_next_locations with batches of batch_size> 1, when setting evaluator_type='local_penalization' but there seems to be a problem with pred (see below) being a list where it should probably be a numpy array.

Doing:

x = np.linspace(-1, 1)
x = x.reshape([len(x),1])

y = x ** 2

domain = [{'name': 'x', 'type':'continuous', 'domain':(-1., 1.)}]

bo = BayesianOptimization(f=None, domain=domain, evaluator_type='local_penalization', batch_size=2, 
                          X=x, Y=y, model_type='GP_MCMC', acquisition_type='EI_MCMC')
bo.suggest_next_locations()

I get error message:
'<' not supported between instances of 'list' and 'float'

And more in detail

~/Work/Envs/develop/lib/python3.6/site-packages/GPyOpt/acquisitions/LP.py in _hammer_function_precompute(self, x0, L, Min, model)
     54         m = model.predict(x0)[0]
     55         pred = model.predict(x0)[1].copy()
---> 56         pred[pred<1e-16] = 1e-16
     57         s = np.sqrt(pred)
     58         r_x0 = (m-Min)/L

Locally I tried pred= np.array(pred), and this solved for batch_size=2, but had a different error message for batch=3. I will create a subsequent issue for this.

When using evaluator_type='random' the original code works fine.

The evaluator_type='thompson_sampling' option yields a different error that I will address in a subsequent issue.

[elzurdo]

I now notice that in my system that the GPyOpt_parallel_optimization.ipynb tutorial does work when I use BO_demo_parallel.suggest_next_locations().

I find that the problem (or my conceptual mistake?) is that I setmodel_type='GP_MCMC' and acquisition_type='EI_MCMC'. When using GP and EI (or LCB), respectively, it seems to work fine.

Since I am not sure if this is a bug in the code, or rather more likely, in my understanding, I leave it for the powers above to decide if to close or further discuss this issue. @javiergonzalezh

[LarsHH]

+1 I get the same error when using GP_MCMC/EI_MCMC with batch size > 1 and local penalization. GP/EI works fine.