Add support to use ipopt to solve constrained problem (#722)

* add support to use ipopt to solve constrainted problem

* fix CI

redisgn IpoptProb

* add authors

* move constraints into Problem class

* update CI

Author: nychiang

.github/workflows/test_hiopbbpy.yml

@@ -14,19 +14,39 @@ jobs:
         uses: actions/checkout@v4
 
       - name: Set up Python
-        uses: actions/setup-python@v4  # Set up Python environment
+        uses: actions/setup-python@v4
         with:
-          python-version: '3.13'
+          python-version: "3.13"
 
+      - name: Set up Conda
+        uses: conda-incubator/setup-miniconda@v2
+        with:
+          auto-activate-base: true
+          channels: conda-forge
+
+      - name: Install IPOPT via conda
+        run: |
+          conda init bash
+          conda install -c conda-forge cyipopt pkg-config
+          export LD_LIBRARY_PATH=$CONDA_PREFIX/lib:$LD_LIBRARY_PATH
 
-      - name: Install HiOpBBPy and its dependencies
+      - name: Verify cyipopt installation
         run: |
+          source ~/.bashrc
+          conda activate base
+          python -c "import cyipopt; print('cyipopt OK')"
+
+      - name: Install HiOpBBPy
+        run: |
+          source ~/.bashrc
+          conda activate base
+          export SKIP_CYIPOPT=1
           python -m pip install --upgrade pip
           pip install .
 
       - name: Run Tests
         run: |
-          python src/Drivers/hiopbbpy/BODriver.py 10
+          source ~/.bashrc
+          conda activate base
+          python src/Drivers/hiopbbpy/BODriverCI.py 10
         continue-on-error: false
-          
-

setup.py

@@ -6,20 +6,26 @@
 '''
 
 import sys
+import os
 import numpy as np
 from setuptools import setup, find_packages
 
 
+install_requires = ["smt"]
+
+if os.getenv("SKIP_CYIPOPT", "0") != "1":
+    install_requires.append("cyipopt")
+
 metadata = dict(
         name="hiopbbpy",
-        version="0.0.3",
+        version="0.0.4",
         description="HiOp black box optimization (hiopbbpy)",
         author="Tucker hartland et al.",
         author_email="hartland1@llnl.gov",
         license="BSD-3",
         packages=find_packages(where="src"),
         package_dir={"": "src"},
-        install_requires=["smt"],
+        install_requires=install_requires,
         python_requires=">=3.9",
         zip_safe=False,
         url="https://github.com/LLNL/hiop",

src/Drivers/hiopbbpy/BODriver.py src/Drivers/hiopbbpy/BODriverCI.pysrc/Drivers/hiopbbpy/BODriver.py renamed to src/Drivers/hiopbbpy/BODriverCI.py

@@ -70,6 +70,15 @@
          min_obj[prob_type][acq_type] = np.inf
          y_opt[prob_type][acq_type] = np.zeros(num_repeat)
 
+      options = {
+        'acquisition_type': acq_type,
+        'bo_maxiter': 20,
+        'opt_solver': 'SLSQP', #"SLSQP" "IPOPT"
+        'solver_options': {
+           'maxiter': 200
+           }
+      }
+
       print("Problem name: ", problem.name)
       print("Acquisition type: ", acq_type)
 
@@ -84,8 +93,8 @@
          gp_model.train(x_train, y_train)
 
          # Instantiate and run Bayesian Optimization
-         bo = BOAlgorithm(gp_model, x_train, y_train, acquisition_type = acq_type) #EI or LCB
-         bo.optimize(problem)
+         bo = BOAlgorithm(problem, gp_model, x_train, y_train, options = options)
+         bo.optimize()
 
          # Retrieve optimal objec
          y_opt[prob_type][acq_type][n_repeat] = bo.getOptimalObjective()

src/Drivers/hiopbbpy/BODriverEX.py

@@ -0,0 +1,95 @@
+"""
+  Code description:
+     for a 2D example LpNormProblem
+        1) randomly sample training points
+        2) define a Kriging-based Gaussian-process (smt backend)
+           trained on said data
+        3) determine the minimizer via BOAlgorithm
+
+Authors:    Tucker Hartland <hartland1@llnl.gov>
+            Nai-Yuan Chiang <chiang7@llnl.gov>
+"""
+
+import sys
+import os
+import numpy as np
+import matplotlib.pyplot as plt
+import warnings
+warnings.filterwarnings("ignore")
+from LpNormProblem import LpNormProblem
+from hiopbbpy.surrogate_modeling import smtKRG
+from hiopbbpy.opt import BOAlgorithm
+from hiopbbpy.problems import BraninProblem
+
+
+# Get user input for the number of repetitions from command-line arguments
+if len(sys.argv) != 2 or int(sys.argv[1]) < 0:
+    num_repeat = 1
+else:
+    num_repeat = int(sys.argv[1])
+
+### parameters
+n_samples = 5  # number of the initial samples to train GP
+theta = 1.e-2  # hyperparameter for GP kernel
+nx = 2         # dimension of the problem
+xlimits = np.array([[-5, 5], [-5, 5]]) # bounds on optimization variable
+
+prob_type_l = ["LpNorm"]      # ["LpNorm", "Branin"]
+acq_type_l = ["LCB"]          # ["LCB", "EI"]
+
+def con_eq(x):
+  return  x[0] + x[1] - 4
+
+def con_jac_eq(x):
+  return  np.array([1.0, 1.0])
+
+def con_ineq(x):
+  return  x[0] - x[1]
+
+def con_jac_ineq(x):
+  return  np.array([1.0, -1.0])
+
+# only 'trust-constr' method supports vector-valued constraints
+user_constraint = [{'type': 'ineq', 'fun': con_ineq, 'jac': con_jac_ineq},
+                   {'type': 'eq',   'fun': con_eq,   'jac': con_jac_eq}]
+
+retval = 0
+for prob_type in prob_type_l:
+   print()
+   if prob_type == "LpNorm":
+      problem = LpNormProblem(nx, xlimits, constraints=None)
+   else:
+      problem = BraninProblem(constraints=None)
+   problem.set_constraints(user_constraint)
+
+   for acq_type in acq_type_l:
+      print("Problem name: ", problem.name)
+      print("Acquisition type: ", acq_type)
+
+      ### initial training set
+      x_train = problem.sample(n_samples)
+      y_train = problem.evaluate(x_train)
+
+      ### Define the GP surrogate model
+      gp_model = smtKRG(theta, xlimits, nx)
+      gp_model.train(x_train, y_train)
+   
+      options = {
+        'acquisition_type': acq_type,
+        'bo_maxiter': 10,
+        'opt_solver': 'IPOPT', #"SLSQP" "IPOPT"
+        'solver_options': {
+           'max_iter': 200,
+           'print_level': 1
+           }
+      }
+   
+      # Instantiate and run Bayesian Optimization
+      bo = BOAlgorithm(problem, gp_model, x_train, y_train, options = options) #EI or LCB
+      bo.optimize()
+
+sys.exit(retval)
+
+
+
+

src/Drivers/hiopbbpy/LpNormProblem.py

@@ -8,9 +8,9 @@
 from hiopbbpy.problems.problem import Problem
 
 class LpNormProblem(Problem):
-    def __init__(self, ndim, xlimits, p=2.0):
+    def __init__(self, ndim, xlimits, p=2.0, constraints=None):
         name = "LpNormProblem"
-        super().__init__(ndim, xlimits, name=name)
+        super().__init__(ndim, xlimits, name=name, constraints=constraints)
         self.p = p
 
     def _evaluate(self, x):

src/hiopbbpy/opt/__init__.py

@@ -1,10 +1,12 @@
 from .boalgorithm import (BOAlgorithmBase, BOAlgorithm)
 from .acquisition import (acquisition, LCBacquisition, EIacquisition)
+from .optproblem import (IpoptProb)
 
 __all__ = [
         "BOAlgorithmBase"
         "BOAlgorithm"
         "acquisition"
         "LCBacquisition"
         "EIacquisition"
+        "IpoptProb"
         ]

src/hiopbbpy/opt/acquisition.py

@@ -14,28 +14,41 @@ class acquisition(object):
     def __init__(self, gpsurrogate):
         assert isinstance(gpsurrogate, GaussianProcess) # add something here
         self.gpsurrogate = gpsurrogate
+        self.has_gradient = False
 
     # Abstract method to evaluate the acquisition function at x.
     def evaluate(self, x: np.ndarray) -> np.ndarray:
         raise NotImplementedError("Child class of acquisition should implement method evaluate")
 
+    # Abstract method to evaluate the gradient of acquisition function at x.
+    def eval_g(self, x: np.ndarray) -> np.ndarray:
+        raise NotImplementedError("Child class of acquisition should implement method evaluate")
 
 # A subclass of acquisition, implementing the Lower Confidence Bound (LCB) acquisition function.
 class LCBacquisition(acquisition):
     def __init__(self, gpsurrogate, beta=3.0):
         super().__init__(gpsurrogate)
         self.beta = beta
+        self.has_gradient = True
 
     # Method to evaluate the acquisition function at x.
     def evaluate(self, x : np.ndarray) -> np.ndarray:
         mu = self.gpsurrogate.mean(x)
-        sig = self.gpsurrogate.variance(x)
-        return mu - self.beta * np.sqrt(sig)
+        sig2 = self.gpsurrogate.variance(x)
+        return mu - self.beta * np.sqrt(sig2)
+
+    def eval_g(self, x: np.ndarray) -> np.ndarray:
+        mu = self.gpsurrogate.mean(x)
+        sig2 = self.gpsurrogate.variance(x)
+        dsig2_dx = self.gpsurrogate.variance_gradient(x)
+        dmu_dx = self.gpsurrogate.mean_gradient(x)
+        return dmu_dx - 0.5 * self.beta * dsig2_dx / np.sqrt(sig2)
 
 # A subclass of acquisition, implementing the Expected improvement (EI) acquisition function.
 class EIacquisition(acquisition):
     def __init__(self, gpsurrogate):
         super().__init__(gpsurrogate)
+        self.has_gradient = True
 
     # Method to evaluate the acquisition function at x.
     def evaluate(self, x : np.ndarray) -> np.ndarray:        
@@ -47,12 +60,41 @@ def evaluate(self, x : np.ndarray) -> np.ndarray:
 
         retval = []
         if sig.size == 1 and np.abs(sig) > 1e-12:
-            arg0 = (y_min - pred) / sig
-            retval = (y_min - pred) * norm.cdf(arg0) + sig * norm.pdf(arg0)
+            z = (y_min - pred) / sig
+            retval = (y_min - pred) * norm.cdf(z) + sig * norm.pdf(z)
             retval *= -1.
         elif sig.size == 1 and np.abs(sig) <= 1e-12:
             retval = 0.0
         elif sig.size > 1:
-            NotImplementedError("TODO --- Not implemented yet!")
+            raise NotImplementedError("TODO --- Not implemented yet!")
 
         return retval
+
+    def eval_g(self, x: np.ndarray) -> np.ndarray:
+        y_data = self.gpsurrogate.training_y
+        y_min = y_data[np.argmin(y_data[:, 0])]
+
+        mean = self.gpsurrogate.mean(x)
+        sig2 = self.gpsurrogate.variance(x)
+        sig = np.sqrt(sig2)
+
+        grad_EI = None
+        if sig.size == 1 and np.abs(sig) > 1e-12:
+            dmean_dx = self.gpsurrogate.mean_gradient(x)
+            dsig2_dx = self.gpsurrogate.variance_gradient(x)
+            dsig_dx = 0.5 * dsig2_dx / sig
+
+            z = (y_min - mean) / sig
+            ncdf = norm.cdf(z)
+            npdf = norm.pdf(z)
+            EI = (y_min - mean) * ncdf + sig * npdf
+
+            dz_dx = -dmean_dx / sig - (y_min - mean) * dsig_dx / sig**2         
+            grad_EI = -dmean_dx * ncdf + dsig_dx * npdf
+            grad_EI *= -1.
+        elif sig.size == 1 and np.abs(sig) <= 1e-12:
+            grad_EI = 0.0
+        elif sig.size > 1:
+            raise NotImplementedError("TODO --- Not implemented yet!")
+
+        return grad_EI