Merge b8fe090 into 3604abe

doc/_src_docs/applications/ego.rstx

@@ -17,7 +17,7 @@ criterion) that determines what the next query point should be.
 One of the earliest bodies of work on Bayesian optimisation that we are aware 
 of are [2]_ and [3]_. Kushner used Wiener processes for one-dimensional problems.
 Kushner’s decision model was based on maximizing the probability of improvement, and included a 
-parameter that controlled the trade-off between ‘more global’ and ‘more local’ optimization, in 
+parameter that controlled the trade-off between 'more global' and 'more local' optimization, in 
 the same spirit as the Exploration/Exploitation trade-off.
 
 Meanwhile, in the former Soviet Union, Mockus and colleagues developed a multidimensional 
@@ -159,6 +159,10 @@ Beside the Expected Improvement, the implementation here offers two other infill
 Regarding the parallel execution, one can implement specific multiprocessing by deriving the _Evaluator_ interface
 and overriding the default implementation of the _run(fun, x)_ method. The default implementation simply runs _fun(x)_.
 
+Regardless the others parameters, you can specify a mixed surrogate model to make mixed optimization. [8]_
+
+You can penalize already evaluated point in EI via tunneling (not recommanded for high dimensional problems).
+
 
 References
 ----------
@@ -177,6 +181,8 @@ References
 
 .. [7] Roux, E. (2011). Assemblage mécanique: stratégies d'optimisation des procédés et d'identification des comportements mécaniques des matériaux (Doctoral dissertation).
 
+.. [8] E.C. Garrido-Merchan and D. Hernandez-Lobato. Dealing with categorical and integer-valued variables in Bayesian Optimization with Gaussian processes. In:Neurocomputing 380 (2020), pages 20–35. 
+
 Usage
 -----
 

doc/_src_docs/surrogate_models/krg.rstx

@@ -27,11 +27,36 @@ More details about the kriging approach could be found in [1]_.
 
 .. [1] Sacks, J. and Schiller, S. B. and Welch, W. J., Designs for computer experiments, Technometrics 31 (1) (1989) 41--47.
 
+
+Kriging with mixed variables 
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+The goal is to be able to build a model for mixed variables. 
+This algorithm has been presented by  Garrido-Merchán and Hernández-Lobato in 2020 [2]_.
+
+To incorporate integer (with order relation) and categorical variables (with no order), we used continuous relaxation.
+For integer, we add a continuous dimension with the same bounds and then we round in the prediction to the closer integer.
+For categorical, we add as many continuous dimensions with bounds [0,1] as possible output values for the variable and then we round in the prediction to the output dimension giving the greatest continuous prediction.
+
+More details avalaible here : 
+.. [2] E. C. Garrido-Merchán and D. Hernández-Lobato. “Dealing with categorical and integer-valued variables in Bayesian Optimization with Gaussian processes”. In:Neurocomputing380 (2020), pages 20–35.doi:10.1016/j.neucom.2019.11.004
+
+
+Implementation Note
+-------------------
+
+The mixed variables are avalaible for all kriging models like KPLS or KPLSK but not to use with gradient.
+
 Usage
 -----
 
 .. embed-test-print-plot :: smt.surrogate_models.tests.test_surrogate_model_examples , Test , test_krg , 80
 
+Usage with parallel options
+^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+.. embed-test-print-plot :: smt.surrogate_models.tests.test_surrogate_model_examples , Test , test_krg_mixed , 80
+
 Options
 -------