AttributeError: 'Param_Discrete_Numeric' object has no attribute 'search_categories'

[dhristozov]

Hi,

Thanks for the nice package.

I am encountering issues when trying to use Param_Discrete_Numeric.

If I understand the code correctly the idea is to use this as a continuous variable during optimisation and mapping the suggested evaluation values to the closest of the numerical categories (via the unit_demap method).

However, by the virtue of Param_Discrete_Numeric inheriting from Param_Discrete this seems to be broken and I get the following exception when trying to use Param_Discrete_Numeric.

params = [
    Param_Categorical("Category", ["Cat-1", "Cat-2", "Cat-3"]),
    Param_Discrete_Numeric("Temperature", list(range(25, 86, 5))),
]
X_space = ParamSpace(params)
target = [
    Target('Desired', aim='max'),
    Target('Undesired', aim='min')
]
campaign = Campaign(X_space, target, seed=42)
X0 = campaign.designer.initialize(4, 'LHS')
Z0 = pd.concat([X0, pd.Series([35.,56.,67.,23.], name="Desired"), pd.Series([60.,48.,27.,70.], name="Undesired")], axis=1)
campaign.add_data(Z0)
campaign.fit()
X_suggest, eval_suggest = campaign.optimizer.suggest(
    acquisition = ['NEHVI', ], m_batch=4
)

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
Cell In[110], [line 15](vscode-notebook-cell:?execution_count=110&line=15)
     [13](vscode-notebook-cell:?execution_count=110&line=13) campaign.add_data(Z0)
     [14](vscode-notebook-cell:?execution_count=110&line=14) campaign.fit()
---> [15](vscode-notebook-cell:?execution_count=110&line=15) X_suggest, eval_suggest = campaign.optimizer.suggest(
     [16](vscode-notebook-cell:?execution_count=110&line=16)     acquisition = ['NEHVI', ], m_batch=4
     [17](vscode-notebook-cell:?execution_count=110&line=17) )

File [~/dev/obsidian/obsidian/optimizer/bayesian.py:711](~/dev/obsidian/obsidian/optimizer/bayesian.py:711), in BayesianOptimizer.suggest(self, m_batch, target, acquisition, optim_sequential, optim_samples, optim_restarts, objective, out_constraints, eq_constraints, ineq_constraints, nleq_constraints, task_index, fixed_var, X_pending, eval_pending)
    [708](~/dev/obsidian/obsidian/optimizer/bayesian.py:708)     raise TypeError('Each item in acquisition list must be either a string or a dictionary')
    [710](~/dev/obsidian/obsidian/optimizer/bayesian.py:710) # Compute static variable inputs
--> [711](~/dev/obsidian/obsidian/optimizer/bayesian.py:711) fixed_features_list = self._fixed_features(fixed_var)
    [713](~/dev/obsidian/obsidian/optimizer/bayesian.py:713) # Set up the sampler, for MC-based optimization of acquisition functions
    [714](~/dev/obsidian/obsidian/optimizer/bayesian.py:714) if not isinstance(model, ModelListGP):

File [~/dev/obsidian/obsidian/optimizer/base.py:114](~/dev/obsidian/obsidian/optimizer/base.py:114), in Optimizer._fixed_features(self, fixed_var)
    [112](~/dev/obsidian/obsidian/optimizer/base.py:112) for x in self.X_space.X_discrete:
    [113](~/dev/obsidian/obsidian/optimizer/base.py:113)     if x.name not in fixed_var.keys():  # Fixed_var should take precedent and lock out other combinations
--> [114](~/dev/obsidian/obsidian/optimizer/base.py:114)         df_i = pd.DataFrame({x.name: x.search_categories})
    [115](~/dev/obsidian/obsidian/optimizer/base.py:115)         df_list.append(df_i)
    [117](~/dev/obsidian/obsidian/optimizer/base.py:117) # Merge by cross

AttributeError: 'Param_Discrete_Numeric' object has no attribute 'search_categories'

[xuyuting]

@dhristozov Hi, thank you for identifying this issue! Could you try the updated code (main branch) again and let us know if it works now?

[dhristozov]

@xuyuting , thanks for the quick fix. I can confirm the exception is gone now. However, looking at the code, I am not sure I understand the purpose of the Param_Discrete_Numeric class. Now (with the code above) I get a warning:

The combinations of discrete features is large at 39. Optimization will proceed very slowly due to the combinatorial explosion. Recommend reducing the number of discrete parameters used.

Which means that my Param_Discrete_Numeric is treated as a categorical variable by the optimiser. If that's the case, may you please explian what is the difference between it and Param_Categorical? E.g., how is Param_Categorical("category", ["1", "2"]) different from Param_Discrete_Numeric("num_category", [1,2]) when it comes to surrogate model building and acquisition optimisation? Thanks!

[fima5]

@dhristozov

Hi! Responding to this comment from a long time ago -

You are getting this error:

The combinations of discrete features is large at 39. Optimization will proceed very slowly due to the combinatorial explosion. Recommend reducing the number of discrete parameters used.

because the optimizer has to enumerate each combination of Param_Discrete_Numeric and Param_Categorical. You have 3 Param_Categorical values and 13 Param_Discrete_Numeric, hence 39 combinations. This is really slow for the optimizer to handle.

As far as the difference between Param_Categorical and Param_Discrete_Numeric: these classes are treated fundamentally different in the optimization function:

• Param_Discrete_Numeric is treated as ordered numeric points, and the optimizer takes into account their numeric relationship (e.g. 25C is closer to 30C than 80C) when building the surrogate model.
• Param_Categorical is treated as an unordered list of values, independent and modeled separately.

Also, the acquisition function you chose: NEHVI, is not known for handling large enums of categorical value pairs well. It's best for continuous or low-cardinality discrete spaces. I would recommend treating temperature as a continuous numerical parameter if possible!