def ei_evaluate_constraint_only(obj_model, constraint_models, cand, current_best, compute_grad):
improvement = lambda mean: np.maximum(0.0, current_best-mean)
improvement_grad = lambda mean_grad: -mean_grad*np.less(mean, current_best).flatten()
# If unconstrained, just compute the GP mean
if len(constraint_models) == 0:
if compute_grad:
mean, var, grad_mean_x, grad_var_x = obj_model.predict(cand, compute_grad=True)
return improvement(mean), improvement_grad(grad_mean_x)
else:
mean, var = obj_model.predict(cand, compute_grad=False)
return improvement(mean)
if cand.ndim == 1:
cand = cand[None]
N_cand = cand.shape[0]
############## ---------------------------------------- ############
############## ############
############## Part that depends on the objective ############
############## ############
############## ---------------------------------------- ############
if current_best is None:
ei = 1.
ei_grad = 0.
else:
target = current_best
# Compute the predictive mean and variance
if not compute_grad:
mean, var = obj_model.predict(cand, compute_grad=False)
else:
mean, var, grad_mean_x, grad_var_x = obj_model.predict(cand, compute_grad=True)
ei_grad = improvement_grad(grad_mean_x)
ei = improvement(mean)
############## ---------------------------------------- ############
############## ############
############## Part that depends on the constraints ############
############## ############
############## ---------------------------------------- ############
# Compute p(valid) for ALL constraints
if not compute_grad:
p_valid_prod = total_constraint_confidence(constraint_models, cand, compute_grad=False)
else:
p_valid_prod, p_grad_prod = total_constraint_confidence(constraint_models, cand, compute_grad=True)
############## ---------------------------------------- ############
############## ############
############## Combine the two parts (obj and con) ############
############## ############
############## ---------------------------------------- ############
acq = ei * p_valid_prod
if not compute_grad:
return acq
else:
return acq, ei_grad * p_valid_prod + p_grad_prod * ei
def constraint_weighted_ei(obj_model, constraint_models, cand, current_best,
compute_grad):
numConstraints = len(constraint_models)
if cand.ndim == 1:
cand = cand[None]
N_cand = cand.shape[0]
############## ---------------------------------------- ############
############## ############
############## Part that depends on the objective ############
############## ############
############## ---------------------------------------- ############
if current_best is None:
ei = 1.
ei_grad = 0.
else:
target = current_best
# Compute the predictive mean and variance
if not compute_grad:
ei = compute_ei(obj_model, cand, target, compute_grad=compute_grad)
else:
ei, ei_grad = compute_ei(obj_model,
cand,
target,
compute_grad=compute_grad)
############## ---------------------------------------- ############
############## ############
############## Part that depends on the constraints ############
############## ############
############## ---------------------------------------- ############
# Compute p(valid) for ALL constraints
if not compute_grad:
p_valid_prod = total_constraint_confidence(constraint_models,
cand,
compute_grad=False)
else:
p_valid_prod, p_grad_prod = total_constraint_confidence(
constraint_models, cand, compute_grad=True)
############## ---------------------------------------- ############
############## ############
############## Combine the two parts (obj and con) ############
############## ############
############## ---------------------------------------- ############
acq = ei * p_valid_prod
if not compute_grad:
return acq
else:
return acq, ei_grad * p_valid_prod + p_grad_prod * ei