Skip to content

Commit

Permalink
Merge pull request #356 from GabrielSoto-INL/BayesianOpt_tempdriverFix
Browse files Browse the repository at this point in the history 
Bayesian Opt input structure fix
  • Loading branch information
@PaulTalbot-INL
PaulTalbot-INL authored Apr 11, 2024
2 parents 2a48899 + 26ffb38 commit febe0ba
Show file tree
Hide file tree
Showing 10 changed files with 201 additions and 88 deletions.
167 changes: 124 additions & 43 deletions src/Cases.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -480,63 +480,118 @@ def _specs_optimizer(cls):
optimizer.addSub(type_sub)

#== Optimization Algorithm ==#
strategy_options = InputTypes.makeEnumType("algorithm", "algorithmType", ['BayesianOpt','GradientDescent'])
strategy_sub = InputData.parameterInputFactory('algorithm', contentType=strategy_options,
descr=r"""Determines what RAVEN optimization algorithm solves the standard HERON TEA.
default is BayesianOpt.""", default='BayesianOpt')
optimizer.addSub(strategy_sub)

#== Persistence ==#
persistenceSub = InputData.parameterInputFactory('persistence',contentType=InputTypes.IntegerType,
descr=r"""provides the number of consecutive times convergence should be reached before a trajectory
is considered fully converged. This helps in preventing early false convergence.""" )
optimizer.addSub(persistenceSub)

#== Optimizer Job Limit ==#
limit_sub = InputData.parameterInputFactory('limit', contentType=InputTypes.IntegerType,
descr=r"""Determines the maximum number of jobs that can submit
for a HERON run. If the optimizer does not converge before this number
of runs, the simulation will terminate. This is helpful to prevent
timeouts.""")
optimizer.addSub(limit_sub)

#== Kernel for BO ==#
strategy_sub = InputData.parameterInputFactory('algorithm', ordered=False,
descr= r"""node detailing which RAVEN optimization
algorithm will solve the standard HERON TEA.""",
default='BayesianOpt')
#====== GradientDescent =======#
gradient_descent = InputData.parameterInputFactory('GradientDescent', ordered=False,
descr= r"""Gradient Descent algorithm for solving
the outer optimization with RAVEN. Settings
specific to the GradientDescent algorithm can
be added within""")
#======== step size factors for GD ========#
growthFactor = InputData.parameterInputFactory('growthFactor', contentType=InputTypes.FloatType,
descr=r"""specifies the rate at which the step size should
grow if the gradient continues in same direction through
multiple iterative steps. For example, a growth factor
of 2 means that if the gradient is identical twice, the
step size is doubled..""", default='2.0')
shrinkFactor = InputData.parameterInputFactory('shrinkFactor', contentType=InputTypes.FloatType,
descr=r"""specifies the rate at which the step size
should shrink if the gradient changes direction through multiple
iterative steps. For example, a shrink factor of 2 means
that if the gradient completely flips direction, the step size is
halved. Note that for stochastic surfaces or low-order gradient
approximations such as SPSA, a small value for the shrink factor
is recommended. If an optimization path appears to be converging
early, increasing the shrink factor might improve the
search.""", default='1.5')
initStepScale = InputData.parameterInputFactory('initialStepScale', contentType=InputTypes.FloatType,
descr=r"""specifies the scale of the initial step in the optimization,
in percent of the size of the problem. The size of the problem
is defined as the hyperdiagonal of the input space, composed of the
input variables. A value of 1 indicates the first step can reach
from the lowest value of all inputs to the highest point of all
inputs, which is too large for all problems with more than one
optimization variable. In general this should be smaller as the number
of optimization variables increases, but large enough that the first
step is meaningful for the problem. This scaling factor should always
be less than $1/\sqrt{N}$, where $N$ is the number of optimization
variables.""", default='0.2')
gradient_descent.addSub(growthFactor)
gradient_descent.addSub(shrinkFactor)
gradient_descent.addSub(initStepScale)
strategy_sub.addSub(gradient_descent)

#====== BayesianOpt =======#
bayesian_opt = InputData.parameterInputFactory('BayesianOpt', ordered=False,
descr= r"""Bayesian Optimization algorithm for solving
the outer optimization with RAVEN. Settings
specific to the BayesianOpt algorithm can
be added within""")
#======== Kernel for BO========#
kernelSub = InputData.parameterInputFactory('kernel', contentType=InputTypes.StringType,
descr=r"""Defines custom kernel expression for GPR used in
Bayesian Optimization. If Gradient Descent is used, then this
node is ignored. Default is Constant*Matern""")
optimizer.addSub(kernelSub)
node is ignored.""",
default='Constant*Matern')
bayesian_opt.addSub(kernelSub)

#== Acquisition Function ==#
#======== Acquisition Function ========#
acquisitionSub = InputData.parameterInputFactory('acquisition', contentType=InputTypes.makeEnumType("acquisition", "acquisitionType",
['ExpectedImprovement', 'ProbabilityOfImprovement','LowerConfidenceBound']),
descr=r"""Node for selecting which acquisition function to use for the
Bayesian Optimizer. If Gradient Descent is used then this node is ignored.
Default is ExpectedImprovement.""")
optimizer.addSub(acquisitionSub)
Bayesian Optimizer.""", default='ExpectedImprovement')
bayesian_opt.addSub(acquisitionSub)

#== Initial Sample Size ==#
initialCountSub = InputData.parameterInputFactory('initialCount', contentType=InputTypes.IntegerType,
descr=r"""Determines number of initial samples for Bayesian Optimization
and number of trajectories for Gradient Descent.""")
optimizer.addSub(initialCountSub)
#======== Kernel for BO========#
seedSub = InputData.parameterInputFactory('seed', contentType=InputTypes.IntegerType,
descr=r"""The seed for random number generator used
for sampling the model within the Bayesian Optimizer.""",
default='RAVEN-determined')
bayesian_opt.addSub(seedSub)

#== Model Selection ==#
#======== Model Selection ========#
modelSelection = InputData.parameterInputFactory('modelSelection',
descr=r"""Parent node for selecting details about how to
update the GPR during optimization for BO.""")
modelDuration = InputData.parameterInputFactory('duration', contentType=InputTypes.IntegerType,
descr=r"""Number of iterations between hyperparameter
selections. Default is 1""")
selections.""", default='1')
modelMethod = InputData.parameterInputFactory('method', contentType=InputTypes.makeEnumType("acquisition", "acquisitionType",
['Internal','External','Average']),
descr=r"""Determines method for selecting hyperparameters.
Internal uses the optimizer within RAVEN to maximize LML. External
uses Scikit-Learn's selection. Averaging samples several models
with MC and then takes a weighted sum. Default is Internal.""")
with MC and then takes a weighted sum.""",
default='Internal')
modelSelection.addSub(modelDuration)
modelSelection.addSub(modelMethod)
optimizer.addSub(modelSelection)
bayesian_opt.addSub(modelSelection)
strategy_sub.addSub(bayesian_opt)
optimizer.addSub(strategy_sub)


#== Persistence ==#
persistenceSub = InputData.parameterInputFactory('persistence',contentType=InputTypes.IntegerType,
descr=r"""provides the number of consecutive times convergence should be reached before a trajectory
is considered fully converged. This helps in preventing early false convergence.""" )
optimizer.addSub(persistenceSub)

#== Optimizer Job Limit ==#
limit_sub = InputData.parameterInputFactory('limit', contentType=InputTypes.IntegerType,
descr=r"""Determines the maximum number of jobs that can submit
for a HERON run. If the optimizer does not converge before this number
of runs, the simulation will terminate. This is helpful to prevent
timeouts.""")
optimizer.addSub(limit_sub)

#== Initial Sample Size ==#
initialCountSub = InputData.parameterInputFactory('initialCount', contentType=InputTypes.IntegerType,
descr=r"""Determines number of initial samples for Bayesian Optimization
and number of trajectories for Gradient Descent.""")
optimizer.addSub(initialCountSub)

#== Convergence Sub Node ==#
convergence = InputData.parameterInputFactory('convergence',
Expand Down Expand Up @@ -828,7 +883,7 @@ def _read_optimization_settings(self, node):
"""
opt_settings = {}

# check first for an opt metric, if not there, default to NPV
# 1. check first for an opt metric, if not there, default to NPV
if node.findFirst('opt_metric') is None:
opt_settings['opt_metric'] = self._default_econ_metric
elif node.findFirst('opt_metric').value == 'LC':
Expand All @@ -837,13 +892,39 @@ def _read_optimization_settings(self, node):
except KeyError:
opt_settings['npv_target'] = 0

for sub in node.subparts:
# 2. check next for algorithm/strategy, need to handle this before "<convergence>"
subparts = node.subparts
algo_head_node = node.findFirst('algorithm')

if algo_head_node is not None:
opt_settings['algorithm'] = {}
if algo_head_node.subparts:
algo = algo_head_node.subparts[0]
algo_dict = {}
# handling case where user submits multiple algorithms, only take the first provided
# when 0 algorithms provided, resorts to defaults. wrong names handled by InputData
if len(algo_head_node.subparts) > 1:
self.raiseAWarning('Multiple optimization algorithms found, only first algorithm ' +
f'provided--{algo.getName()}--will be used.')
# handle extra settings to override in the outer.xml
for sub_algo in algo.subparts:
sub_name = sub_algo.getName()
if sub_name == 'modelSelection':
algo_dict[sub_name] = {}
for ssub in sub_algo.subparts:
algo_dict[sub_name][ssub.getName()] = ssub.value
else:
algo_dict[sub_name] = sub_algo.value
# save back to optimization_settings
opt_settings['algorithm'][algo.getName()] = algo_dict
self._optimization_strategy = algo.getName()
subparts.remove(algo_head_node)

# 3. handle rest of the inputs in the optimization_settings node
for sub in subparts:
sub_name = sub.getName()
# optimization algorithm
if sub_name == 'algorithm':
self._optimization_strategy = sub.value
# add metric information to opt_settings dictionary
elif sub_name == 'stats_metric':
if sub_name == 'stats_metric':
opt_settings[sub_name] = {}
stats_metric_name = sub.value
opt_settings[sub_name]['name'] = stats_metric_name
Expand All @@ -863,7 +944,7 @@ def _read_optimization_settings(self, node):
opt_settings[sub_name]['threshold'] = sub.parameterValues['threshold']
except KeyError:
opt_settings[sub_name]['threshold'] = 0.05
elif sub_name in ['convergence', 'modelSelection']:
elif sub_name == 'convergence':
opt_settings[sub_name] = {}
for ssub in sub.subparts:
if ssub.getName() == 'objective' and self._optimization_strategy == 'BayesianOpt':
Expand Down
2 changes: 1 addition & 1 deletion templates/outer.xml
View file
Original file line number Diff line number Diff line change
Expand Up @@ -65,7 +65,7 @@
</Code>
<ROM name="gpROM" subType="GaussianProcessRegressor">
<Features></Features>
<Target></Target>
<Target>mean_NPV</Target>
<alpha>1e-8</alpha>
<n_restarts_optimizer>5</n_restarts_optimizer>
<normalize_y>True</normalize_y>
Expand Down
Loading

0 comments on commit febe0ba

Please sign in to comment.