Pr/bo only

docs/algos_explained.rst

@@ -30,3 +30,4 @@ in ``Propulate``.
    pso
    cmaes
    nm
+   bayesian_optimizer

docs/autoapi/index.rst

@@ -0,0 +1,11 @@
+API Reference
+=============
+
+This page contains auto-generated API reference documentation [#f1]_.
+
+.. toctree::
+   :titlesonly:
+
+   /autoapi/propulate/index
+
+.. [#f1] Created with `sphinx-autoapi <https://github.com/readthedocs/sphinx-autoapi>`_

docs/tut_bayesopt.rst

@@ -0,0 +1,31 @@
+.. _tut_bayesopt:
+
+Bayesian Optimizer Tutorial
+===========================
+
+This page links the practical BO usage material in ``Propulate``:
+
+- Theory and implementation details: :ref:`Bayesian Optimization page <bayesopt>`
+- Runnable comparison script: ``tutorials/bayesian_optimizer_example.py``
+
+Quick run with default EI acquisition from repository root:
+
+.. code-block:: console
+
+   mpirun --use-hwthread-cpus -n 4 python tutorials/bayesian_optimizer_example.py \
+     --function sphere \
+     --generations 25 \
+     --checkpoint ./bo_runs
+
+Run with Thompson Sampling instead (no hyperparameters to tune; diversity is automatic):
+
+.. code-block:: console
+
+   mpirun --use-hwthread-cpus -n 4 python tutorials/bayesian_optimizer_example.py \
+     --function sphere \
+     --generations 25 \
+     --bo-acq TS \
+     --checkpoint ./bo_runs_ts
+
+For minimal custom setups with ``BayesianOptimizer`` + ``Propulator``, including
+Thompson Sampling and mixed-type examples, see :ref:`the BO documentation page <bayesopt>`.

docs/tut_propulator.rst

@@ -258,6 +258,12 @@ optimizer flavors are available, along with example scripts showing how to use t
   works. Check out the example script for how to use Nelder-Mead in ``Propulate`` |:dna:|:
   https://github.com/Helmholtz-AI-Energy/propulate/blob/master/tutorials/nm_example.py
 
+**Bayesian Optimization (BO)**
+  Uses a Gaussian-process surrogate and an acquisition function to propose sample-efficient evaluations for expensive
+  black-box objectives. :ref:`Here<bayesopt>` you can find a more detailed explanation of the implementation in
+  ``Propulate`` |:dna:|. Check out the example script:
+  https://github.com/Helmholtz-AI-Energy/propulate/blob/master/tutorials/bayesian_optimizer_example.py
+
 
 [1] *N. Hansen and A. Ostermeier (2001), "Completely Derandomized Self-Adaptation in Evolution Strategies", Evolutionary Computation, 9(2), 159-195.*
 https://doi.org/10.1162/106365601750190398

docs/usage.rst

@@ -15,6 +15,7 @@ If you want to use ``Propulate`` for your own applications, you can use these sc
    tut_surrogate
    tut_multi_rank_worker
    tut_ddp
+   tut_bayesopt
 
 .. Links
 .. _Github: https://github.com/Helmholtz-AI-Energy/propulate

propulate/population.py

@@ -15,6 +15,21 @@
 import numpy as np
 
 
+def _normalize_param_type(value: Any) -> type:
+    """Map a limit value to its canonical Python type (float, int, or str).
+
+    Handles numpy scalar types (e.g. ``np.int64``, ``np.float32``) so that
+    downstream ``is`` checks against ``int`` / ``float`` work reliably.
+    """
+    if isinstance(value, str):
+        return str
+    if isinstance(value, (int, np.integer)):
+        return int
+    if isinstance(value, (float, np.floating)):
+        return float
+    raise TypeError(f"Unsupported parameter type: {type(value)}")
+
+
 class Individual:
     """An individual represents a candidate solution to the considered optimization problem."""
 
@@ -42,7 +57,7 @@ def __init__(
             if key.startswith("_"):
                 raise ValueError("Keys starting with '_' are reserved.")
         # NOTE keep track of the types of variables for setting and getting
-        self.types = {key: type(limits[key][0]) for key in limits}
+        self.types = {key: _normalize_param_type(limits[key][0]) for key in limits}
         # NOTE offsets are used to keep track of where each variable is stored in the position field, since a categorical embedding can take up more space than other types of variables
         offset = 0
         self.offsets = {}
@@ -119,8 +134,8 @@ def __setitem__(self, key: str, newvalue: Union[float, int, str, Any]) -> None:
             elif self.types[key] is str:
                 assert newvalue in self.limits[key]
                 offset = self.offsets[key]
-                upper = len(self.limits[key])
-                self.position[offset:upper] = np.array([0])
+                upper = offset + len(self.limits[key])
+                self.position[offset:upper] = 0.0
                 self.position[offset + self.limits[key].index(newvalue)] = 1.0
             else:
                 raise ValueError("Unknown type")

propulate/propagators/__init__.py

@@ -10,6 +10,22 @@
     SelectUniform,
     Stochastic,
 )
+from .bayesopt import (
+    AcquisitionFunction,
+    BayesianOptimizer,
+    ExpectedImprovement,
+    FitterType,
+    LowerConfidenceBound,
+    MultiStartAcquisitionOptimizer,
+    ProbabilityImprovement,
+    SingleCPUFitter,
+    SurrogateFitter,
+    UpperConfidenceBound,
+    create_acquisition,
+    create_fitter,
+    expected_improvement,
+    get_default_kernel_sklearn,
+)
 from .cmaes import (
     ActiveCMA,
     BasicCMA,
@@ -60,4 +76,18 @@
     "BasicCMA",
     "ActiveCMA",
     "ParallelNelderMead",
+    "BayesianOptimizer",
+    "expected_improvement",
+    "create_acquisition",
+    "AcquisitionFunction",
+    "ExpectedImprovement",
+    "ProbabilityImprovement",
+    "UpperConfidenceBound",
+    "LowerConfidenceBound",
+    "MultiStartAcquisitionOptimizer",
+    "SurrogateFitter",
+    "FitterType",
+    "SingleCPUFitter",
+    "create_fitter",
+    "get_default_kernel_sklearn",
 ]

propulate/propagators/base.py

@@ -44,6 +44,8 @@ class Propagator:
     -------
     __call__()
         Apply the propagator.
+    set_worker_context()
+        Configure worker-rank context for rank-sensitive propagators.
     """
 
     def __init__(self, parents: int = 0, offspring: int = 0, rng: Optional[random.Random] = None) -> None:
@@ -95,6 +97,23 @@ def __call__(self, inds: List[Individual]) -> Union[List[Individual], Individual
         """
         raise NotImplementedError()
 
+    def set_worker_context(self, worker_rank: int, worker_size: int) -> None:
+        """Set worker communicator context for rank-sensitive propagators.
+
+        Parameters
+        ----------
+        worker_rank : int
+            Rank in the worker/island communicator.
+        worker_size : int
+            Number of workers in the worker/island communicator.
+
+        Notes
+        -----
+        Default is a no-op to keep backward compatibility. Propagators that
+        depend on rank-local state (e.g. BO/PSO) can override this method.
+        """
+        _ = (worker_rank, worker_size)
+
 
 class Stochastic(Propagator):
     """
@@ -221,6 +240,11 @@ def __call__(self, inds: List[Individual]) -> Union[List[Individual], Individual
         else:  # Else apply `false_prop`.
             return self.false_prop(inds)
 
+    def set_worker_context(self, worker_rank: int, worker_size: int) -> None:
+        """Propagate worker context to wrapped propagators."""
+        self.true_prop.set_worker_context(worker_rank, worker_size)
+        self.false_prop.set_worker_context(worker_rank, worker_size)
+
 
 class Compose(Propagator):
     """
@@ -288,6 +312,11 @@ def __call__(self, inds: List[Individual]) -> Union[List[Individual], Individual
             inds = p(inds)  # type: ignore
         return inds
 
+    def set_worker_context(self, worker_rank: int, worker_size: int) -> None:
+        """Propagate worker context to all composed propagators."""
+        for p in self.propagators:
+            p.set_worker_context(worker_rank, worker_size)
+
 
 class SelectMin(Propagator):
     """