Dimensionality issues of the gradient with linear-multi-fidelity model when evaluating Expected improvement with context variables

[ToennisStef]

Hello,
I want to do a Bayesian optimization of a 2-dimensional multi-fidelity test problem. I therefore defined a linear multi-fidelity model and set up the acquisition function. I also set a context variable for the fidelity parameter. I set the fidelity parameter value for the optimization to the highest fidelity. But when trying to optimize the acquisition function an error returned immediately, originating form the scipy l-bfgs algorithm. I noticed that the shape of the gradient of the acquisition function was (1,2,2) instead of (2) or (1,2). I noticed the dimensionality error came from the expected improvement acquisition function definition specificially from this line:

dimprovement_dx = dstandard_deviation_dx * pdf - cdf * dmean_dx

the shape of dmean_dx was (1,2,1), this results in the dimprovement_dx.shape being (1,2,2).
quick fix for the evaluation was to add this line of code:

dimprovement_dx = np.diagonal(dimprovement_dx).T

which worked now for my 2-dimensional test problem. It will hopefully also still work for 1-dimensional test problems, but i think for higher dimensional problems it will not work, so its probably advisable to just fix the dimensionality of dmean_dx.

Code to replicat the error is here:

import numpy as np
import GPy

from emukit.core import ParameterSpace, ContinuousParameter, DiscreteParameter
from emukit.bayesian_optimization.acquisitions import ExpectedImprovement
from emukit.core.optimization import GradientAcquisitionOptimizer
import emukit.multi_fidelity
from emukit.model_wrappers.gpy_model_wrappers import GPyMultiOutputWrapper
from emukit.multi_fidelity.models import GPyLinearMultiFidelityModel


X_train = np.array([[ 0.13473929, -0.32606444,  0.        ],
       [-1.53102551,  1.4245339 ,  0.        ],
       [-0.48998839, -1.29864283,  0.        ],
       [ 1.84718222,  0.45174302,  0.        ],
       [ 1.32512917, -1.85203682,  0.        ],
       [-0.96369396,  0.89760785,  0.        ],
       [-1.04751664, -0.76983995,  0.        ],
       [ 0.66267413,  1.9795312 ,  0.        ],
       [ 0.80437702, -1.13627663,  0.        ],
       [-1.40778594,  0.11410064,  0.        ],
       [-0.5741273 , -0.24182373,  0.        ],
       [ 1.21662989,  1.00876794,  0.        ],
       [ 1.73703451, -0.66999665,  0.        ],
       [-0.0981639 ,  1.57938117,  0.        ],
       [-1.88964734, -1.70531898,  0.        ],
       [ 0.27418341,  0.54433427,  0.        ],
       [ 0.47281786, -1.60549965,  0.        ],
       [-1.86141352,  0.64387054,  0.        ],
       [-0.14885246, -0.51576477,  0.        ],
       [ 1.51398075,  1.73388086,  0.        ],
       [-1.43824131,  0.65202268,  1.        ],
       [ 0.12916456, -1.85884469,  1.        ],
       [ 1.68246282,  1.32535729,  1.        ],
       [-0.99837652, -0.18159998,  1.        ],
       [-0.21917494,  1.69523472,  1.        ],
       [ 1.41106617, -0.80000571,  1.        ],
       [ 0.90085879,  0.26700687,  1.        ],
       [-1.71775573, -1.22432354,  1.        ],
       [-1.8063743 ,  1.13146007,  1.        ],
       [ 0.62229842, -0.36158118,  1.        ],
       [ 1.06610087,  0.82823509,  1.        ],
       [-0.25701166, -1.66090378,  1.        ],
       [-0.52450227,  0.21373441,  1.        ],
       [ 1.84139629, -1.29492998,  1.        ],
       [ 0.34669023,  1.76207188,  1.        ],
       [-1.03858617, -0.74269003,  1.        ],
       [-1.24743496,  1.88256894,  1.        ],
       [ 0.43151265, -0.61242308,  1.        ],
       [ 1.88211933,  0.0793515 ,  1.        ],
       [-0.69120659, -1.41173808,  1.        ],
       [ 1.75531639,  0.01006163,  2.        ],
       [-0.16598199, -1.41704852,  2.        ],
       [-1.39763355,  1.91486136,  2.        ],
       [ 0.55827226, -0.50934682,  2.        ],
       [ 0.28320662,  1.19174326,  2.        ],
       [-1.63918466, -0.34873588,  2.        ],
       [-0.92222783,  0.8503596 ,  2.        ],
       [ 1.02817909, -1.6948546 ,  2.        ],
       [ 1.41036964,  1.53844309,  2.        ],
       [-0.50848947, -0.88097757,  2.        ],
       [-1.80314409,  0.38062028,  2.        ],
       [ 0.15129807, -1.03565238,  2.        ],
       [ 0.87819937,  0.72781515,  2.        ],
       [-1.04663165, -1.82035536,  2.        ],
       [-0.2667013 ,  1.32405309,  2.        ],
       [ 1.68516784, -0.22909473,  2.        ],
       [ 1.51735547,  1.44219314,  2.        ],
       [-0.43482649, -0.09917109,  2.        ],
       [-1.12851415,  0.59992489,  2.        ],
       [ 0.79601932, -1.94440427,  2.        ]])

Y_train = np.array([[1.24180338e+00],
       [9.84373290e+00],
       [2.44074890e+01],
       [8.03973669e+01],
       [1.22073038e+02],
       [4.11819471e-01],
       [3.71351316e+01],
       [2.31060430e+01],
       [3.05597094e+01],
       [3.80808397e+01],
       [3.57512288e+00],
       [2.20424659e+00],
       [1.25195017e+02],
       [2.49581510e+01],
       [3.08139883e+02],
       [2.26366363e+00],
       [3.26536686e+01],
       [8.85821224e+01],
       [3.01479215e+00],
       [3.07271610e+00],
       [1.01034474e+02],
       [1.81278479e+02],
       [1.25356617e+02],
       [7.10193407e+01],
       [1.38096216e+02],
       [4.00845553e+02],
       [2.26572223e+01],
       [8.73085407e+02],
       [2.27545534e+02],
       [3.47838964e+01],
       [1.32073600e+01],
       [1.53116297e+02],
       [2.52078040e+00],
       [1.11225838e+03],
       [1.39240704e+02],
       [1.67680976e+02],
       [5.57808268e+00],
       [3.78929038e+01],
       [6.13716074e+02],
       [1.81275800e+02],
       [9.43720052e+02],
       [2.10046008e+02],
       [5.89701164e+00],
       [6.76016433e+01],
       [1.24065302e+02],
       [9.28489813e+02],
       [3.69496190e+00],
       [7.57354963e+02],
       [2.04814007e+01],
       [1.32130481e+02],
       [8.31954239e+02],
       [1.12771727e+02],
       [2.03356232e-01],
       [8.54373681e+02],
       [1.58586264e+02],
       [9.42275205e+02],
       [7.42576705e+01],
       [1.03672545e+01],
       [4.99068690e+01],
       [6.64676317e+02]])

# Define kernels
kernels = [GPy.kern.Matern52(input_dim=2, ARD=True), GPy.kern.Matern52(input_dim=2, ARD=True), GPy.kern.Matern52(input_dim=2, ARD=True)]
lin_mf_kernel = emukit.multi_fidelity.kernels.LinearMultiFidelityKernel(kernels)

# Define Multi-fidelity model
gpy_lin_mf_model = GPyLinearMultiFidelityModel(X_train, Y_train, lin_mf_kernel, n_fidelities=3)

# Fix noise
gpy_lin_mf_model.mixed_noise.Gaussian_noise.fix(0)
gpy_lin_mf_model.mixed_noise.Gaussian_noise_1.fix(0)
gpy_lin_mf_model.mixed_noise.Gaussian_noise_2.fix(0)

# Wrap the model using the given 'GPyMultiOutputWrapper'
lin_mf_model = model = GPyMultiOutputWrapper(gpy_lin_mf_model, n_outputs=3, n_optimization_restarts=20)

## Fit the model
lin_mf_model.optimize()

# Define parameter space
parameter_space = ParameterSpace([ContinuousParameter('x1', 0, 1), ContinuousParameter('x2', 0, 1), DiscreteParameter('fid', [0, 1, 2])])

# Define acquisition function
ei_acquisition = ExpectedImprovement(model = lin_mf_model)

optimizer = GradientAcquisitionOptimizer(parameter_space, num_anchor=1)

new_x, _  = optimizer.optimize(ei_acquisition, context={'fid': 2})

Sorry for the messy training data. I thought it would be easiest to just copy my current training data.
I hope even my fix was kind of correct for 2-dimensional test problems. I am happy for any input on this matter.
Happy new year and best regards,
Stefan Tönnis

As quick edit:
This is the error message:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
ValueError: too many axes: 2 (effrank=2), expected rank=1


The above exception was the direct cause of the following exception:

ValueError                                Traceback (most recent call last)
Cell In[21], [line 19](vscode-notebook-cell:?execution_count=21&line=19)
     [15](vscode-notebook-cell:?execution_count=21&line=15)             return np.min(self.model.Y[(Problem.n_fidelities-1)*10*Problem.dim:], axis=0)
     [17](vscode-notebook-cell:?execution_count=21&line=17) log_cei_acquisition = LogAcquisition(CustomExpectedImprovement(model = lin_mf_model))
---> [19](vscode-notebook-cell:?execution_count=21&line=19) new_x, _  = optimizer.optimize(Log_ei_acquisition, context={'fid': 2})
     [20](vscode-notebook-cell:?execution_count=21&line=20) new_xc, _ = optimizer.optimize(log_cei_acquisition, context={'fid': 2})
     [22](vscode-notebook-cell:?execution_count=21&line=22) acq_val = log_cei_acquisition.evaluate(X_plot_h)

File c:\Users\StefanT\.conda\envs\ML\Lib\site-packages\emukit\core\optimization\acquisition_optimizer.py:53, in AcquisitionOptimizerBase.optimize(self, acquisition, context)
     [50](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/emukit/core/optimization/acquisition_optimizer.py:50)     context = dict()
     [51](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/emukit/core/optimization/acquisition_optimizer.py:51) context_manager = ContextManager(self.space, context)
---> [53](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/emukit/core/optimization/acquisition_optimizer.py:53) max_x, max_value = self._optimize(acquisition, context_manager)
     [55](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/emukit/core/optimization/acquisition_optimizer.py:55) # Optimization might not match any encoding exactly
     [56](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/emukit/core/optimization/acquisition_optimizer.py:56) # Rounding operation here finds the closest encoding
     [57](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/emukit/core/optimization/acquisition_optimizer.py:57) rounded_max_x = self.space.round(max_x)

File c:\Users\StefanT\.conda\envs\ML\Lib\site-packages\emukit\core\optimization\gradient_acquisition_optimizer.py:72, in GradientAcquisitionOptimizer._optimize(self, acquisition, context_manager)
     [70](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/emukit/core/optimization/gradient_acquisition_optimizer.py:70) optimized_points = []
     [71](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/emukit/core/optimization/gradient_acquisition_optimizer.py:71) for a in anchor_points:
---> [72](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/emukit/core/optimization/gradient_acquisition_optimizer.py:72)     optimized_point = apply_optimizer(
     [73](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/emukit/core/optimization/gradient_acquisition_optimizer.py:73)         optimizer, a, space=self.space, f=f, df=None, f_df=f_df, context_manager=context_manager
     [74](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/emukit/core/optimization/gradient_acquisition_optimizer.py:74)     )
     [75](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/emukit/core/optimization/gradient_acquisition_optimizer.py:75)     optimized_points.append(optimized_point)
     [77](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/emukit/core/optimization/gradient_acquisition_optimizer.py:77) x_min, fx_min = min(optimized_points, key=lambda t: t[1])

File c:\Users\StefanT\.conda\envs\ML\Lib\site-packages\emukit\core\optimization\optimizer.py:134, in apply_optimizer(optimizer, x0, space, f, df, f_df, context_manager)
    [130](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/emukit/core/optimization/optimizer.py:130) else:
    [131](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/emukit/core/optimization/optimizer.py:131)     f_df_no_context = problem.f_df_no_context
--> [134](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/emukit/core/optimization/optimizer.py:134) optimized_x, _ = optimizer.optimize(problem.x0_no_context, f_no_context, df_no_context, f_df_no_context)
    [136](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/emukit/core/optimization/optimizer.py:136) # Add context and round according to the type of variables of the design space
    [137](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/emukit/core/optimization/optimizer.py:137) suggested_x_with_context = add_context(optimized_x)

File c:\Users\StefanT\.conda\envs\ML\Lib\site-packages\emukit\core\optimization\optimizer.py:73, in OptLbfgs.optimize(self, x0, f, df, f_df)
     [69](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/emukit/core/optimization/optimizer.py:69)     res = scipy.optimize.fmin_l_bfgs_b(
     [70](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/emukit/core/optimization/optimizer.py:70)         f, x0=x0, bounds=self.bounds, approx_grad=True, maxiter=self.max_iterations
     [71](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/emukit/core/optimization/optimizer.py:71)     )
     [72](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/emukit/core/optimization/optimizer.py:72) else:
---> [73](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/emukit/core/optimization/optimizer.py:73)     res = scipy.optimize.fmin_l_bfgs_b(_f_df, x0=x0, bounds=self.bounds, maxiter=self.max_iterations)
     [75](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/emukit/core/optimization/optimizer.py:75) # We check here if the the optimizer moved. It it didn't we report x0 and f(x0) as scipy can return NaNs
     [76](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/emukit/core/optimization/optimizer.py:76) if res[2]["task"] == b"ABNORMAL_TERMINATION_IN_LNSRCH":

File c:\Users\StefanT\.conda\envs\ML\Lib\site-packages\scipy\optimize\_lbfgsb_py.py:199, in fmin_l_bfgs_b(func, x0, fprime, args, approx_grad, bounds, m, factr, pgtol, epsilon, iprint, maxfun, maxiter, disp, callback, maxls)
    [187](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/scipy/optimize/_lbfgsb_py.py:187) callback = _wrap_callback(callback)
    [188](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/scipy/optimize/_lbfgsb_py.py:188) opts = {'disp': disp,
    [189](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/scipy/optimize/_lbfgsb_py.py:189)         'iprint': iprint,
    [190](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/scipy/optimize/_lbfgsb_py.py:190)         'maxcor': m,
   (...)
    [196](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/scipy/optimize/_lbfgsb_py.py:196)         'callback': callback,
    [197](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/scipy/optimize/_lbfgsb_py.py:197)         'maxls': maxls}
--> [199](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/scipy/optimize/_lbfgsb_py.py:199) res = _minimize_lbfgsb(fun, x0, args=args, jac=jac, bounds=bounds,
    [200](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/scipy/optimize/_lbfgsb_py.py:200)                        **opts)
    [201](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/scipy/optimize/_lbfgsb_py.py:201) d = {'grad': res['jac'],
    [202](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/scipy/optimize/_lbfgsb_py.py:202)      'task': res['message'],
    [203](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/scipy/optimize/_lbfgsb_py.py:203)      'funcalls': res['nfev'],
    [204](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/scipy/optimize/_lbfgsb_py.py:204)      'nit': res['nit'],
    [205](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/scipy/optimize/_lbfgsb_py.py:205)      'warnflag': res['status']}
    [206](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/scipy/optimize/_lbfgsb_py.py:206) f = res['fun']

File c:\Users\StefanT\.conda\envs\ML\Lib\site-packages\scipy\optimize\_lbfgsb_py.py:360, in _minimize_lbfgsb(fun, x0, args, jac, bounds, disp, maxcor, ftol, gtol, eps, maxfun, maxiter, iprint, callback, maxls, finite_diff_rel_step, **unknown_options)
    [358](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/scipy/optimize/_lbfgsb_py.py:358) g = g.astype(np.float64)
    [359](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/scipy/optimize/_lbfgsb_py.py:359) # x, f, g, wa, iwa, task, csave, lsave, isave, dsave = \
--> [360](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/scipy/optimize/_lbfgsb_py.py:360) _lbfgsb.setulb(m, x, low_bnd, upper_bnd, nbd, f, g, factr,
    [361](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/scipy/optimize/_lbfgsb_py.py:361)                pgtol, wa, iwa, task, iprint, csave, lsave,
    [362](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/scipy/optimize/_lbfgsb_py.py:362)                isave, dsave, maxls)
    [363](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/scipy/optimize/_lbfgsb_py.py:363) task_str = task.tobytes()
    [364](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/scipy/optimize/_lbfgsb_py.py:364) if task_str.startswith(b'FG'):
    [365](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/scipy/optimize/_lbfgsb_py.py:365)     # The minimization routine wants f and g at the current x.
    [366](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/scipy/optimize/_lbfgsb_py.py:366)     # Note that interruptions due to maxfun are postponed
    [367](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/scipy/optimize/_lbfgsb_py.py:367)     # until the completion of the current minimization iteration.
    [368](file:///C:/Users/StefanT/.conda/envs/ML/Lib/site-packages/scipy/optimize/_lbfgsb_py.py:368)     # Overwrite f and g:

ValueError: failed in converting 7th argument `g' of _lbfgsb.setulb to C/Fortran array

The function call is a bit different also i used the log expected improvement but the problem/error is the same.