Split and organize first pass.

This commit is contained in:
James Pace 2022-07-06 12:10:43 +00:00
parent 6ab8166ffc
commit a4498cd33d
8 changed files with 284 additions and 236 deletions

View File

@ -8,14 +8,17 @@ find_package(ament_cmake REQUIRED)
find_package(ament_cmake_ros REQUIRED)
find_package(rclcpp REQUIRED)
add_library(optimizer src/optimizer.cpp)
target_compile_features(optimizer PUBLIC cxx_std_17) # Require C++17
target_include_directories(optimizer PUBLIC
add_library(simplex_solver
src/CostFunction.cpp
src/SimplexSolver.cpp)
target_compile_features(simplex_solver PUBLIC cxx_std_17) # Require C++17
target_include_directories(simplex_solver PUBLIC
$<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>
$<INSTALL_INTERFACE:include>)
add_executable(main src/main.cpp)
target_compile_features(main PUBLIC cxx_std_17) # Require C++17
target_link_libraries(main simplex_solver)
target_include_directories(main PUBLIC
$<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>
$<INSTALL_INTERFACE:include>)
@ -27,7 +30,7 @@ install(
)
install(
TARGETS
optimizer
simplex_solver
EXPORT export_${PROJECT_NAME}
ARCHIVE DESTINATION lib/${PROJECT_NAME}
LIBRARY DESTINATION lib/${PROJECT_NAME}

View File

@ -7,20 +7,25 @@
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#ifndef J7S__OPTIMIZER_HPP_
#define J7S__OPTIMIZER_HPP_
#ifndef J7S__COSTFUNCTION_HPP_
#define J7S__COSTFUNCTION_HPP_
namespace j7s
{
class Optimizer
class CostFunction
{
public:
Optimizer();
CostFunction(double a, double b, double c);
double eval(double input) const;
virtual ~Optimizer();
double actualBest() const;
private:
double m_a;
double m_b;
double m_c;
};
} // namespace j7s
#endif // J7S__OPTIMIZER_HPP_
#endif // J7S__COSTFUNCTION_HPP_

View File

@ -0,0 +1,42 @@
// Copyright 2022 James Pace
// All Rights Reserved.
//
// For a license to this software contact
// James Pace at jpace121@gmail.com.
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#ifndef J7S__SIMPLEXSOLVER_HPP_
#define J7S__SIMPLEXSOLVER_HPP_
#include "j7s-optimization/CostFunction.hpp"
#include "j7s-optimization/common.hpp"
#include <vector>
namespace j7s
{
class SimplexSolver
{
public:
SimplexSolver(const CostFunction & costFunction, const std::vector<double> initSimplex);
IterationState update();
Coordinate bestCoord() const;
private:
const CostFunction m_costFunction;
std::vector<Coordinate> m_currentSimplex;
// Helper functions.
double newPoint();
std::vector<Coordinate> contract();
double calcVolume();
};
} // namespace j7s
#endif // J7S__SIMPLEXSOLVER_HPP_

View File

@ -0,0 +1,41 @@
// Copyright 2022 James Pace
// All Rights Reserved.
//
// For a license to this software contact
// James Pace at jpace121@gmail.com.
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#ifndef J7S__COMMON_HPP_
#define J7S__COMMON_HPP_
namespace j7s
{
// The state after a single iteration of the solver.
enum class IterationState
{
OK,
CONVERGED
};
// A coordinate is pair of a cost input and the resulting the cost.
// We save them together to minimize cost function evaluations.
// This isn't the cleanest thing in the world, but it's worth saving
// evaluations.
struct Coordinate
{
double input;
double cost;
Coordinate() : input{0.0}, cost{0.0} {};
Coordinate(double input, double cost) : input{input}, cost{cost} {};
// Sort by cost.
bool operator<(const Coordinate & other) const { return (cost < other.cost); }
};
} // namespace j7s
#endif // J7S__COMMON_HPP_

33
src/CostFunction.cpp Normal file
View File

@ -0,0 +1,33 @@
// Copyright 2022 James Pace
// All Rights Reserved.
//
// For a license to this software contact
// James Pace at jpace121@gmail.com.
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#include "j7s-optimization/CostFunction.hpp"
#include <cmath>
namespace j7s
{
CostFunction::CostFunction(double a, double b, double c) : m_a{a}, m_b{b}, m_c{c} {}
double CostFunction::eval(double input) const
{
return m_a * std::pow(input, 2) + m_b * input + m_c;
}
double CostFunction::actualBest() const
{
// y = a*x**2 + b*x + c
// dy/dx = 2*a*x + b
// 0 = 2*a*x + b
// -b/2*a = x
return (-m_b / (2 * m_a));
}
} // namespace j7s

138
src/SimplexSolver.cpp Normal file
View File

@ -0,0 +1,138 @@
// Copyright 2022 James Pace
// All Rights Reserved.
//
// For a license to this software contact
// James Pace at jpace121@gmail.com.
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#include "j7s-optimization/SimplexSolver.hpp"
#include <algorithm>
#include <cmath>
#include <stdexcept>
namespace j7s
{
SimplexSolver::SimplexSolver(
const CostFunction & costFunction, const std::vector<double> initSimplex) :
m_costFunction{costFunction}
{
m_currentSimplex.reserve(initSimplex.size());
for (const auto val : initSimplex)
{
m_currentSimplex.emplace_back(val, m_costFunction.eval(val));
}
std::sort(m_currentSimplex.begin(), m_currentSimplex.end());
}
Coordinate SimplexSolver::bestCoord() const
{
return m_currentSimplex.front();
}
double SimplexSolver::newPoint()
{
if (m_currentSimplex.size() == 0)
{
throw std::runtime_error("Simplex can't be missing.");
}
// Calculate sum.
double biggest = m_currentSimplex.back().input;
// Calculate volume.
double sum = 0.0;
// All but the biggest, (would be adding 0...)
for (unsigned int index = 0; index < m_currentSimplex.size() - 1; index++)
{
const double diff = m_currentSimplex[index].input - biggest;
sum += diff;
}
const double newPoint = sum * (2.0 / m_currentSimplex.size());
return newPoint;
}
std::vector<Coordinate> SimplexSolver::contract()
{
const auto smallest = m_currentSimplex.front();
std::vector<Coordinate> newVector;
newVector.reserve(m_currentSimplex.size());
newVector.emplace_back(smallest);
// TODO: Really check size before I get here...
for (auto it = m_currentSimplex.begin() + 1; it != m_currentSimplex.end(); it++)
{
const auto oldInput = it->input;
const auto newInput = 0.5 * (oldInput + smallest.input);
const auto newCost = m_costFunction.eval(newInput);
newVector.emplace_back(newInput, newCost);
}
std::sort(newVector.begin(), newVector.end());
return newVector;
}
double SimplexSolver::calcVolume()
{
// TODO: For reals do something like:
// https://math.stackexchange.com/questions/337197/finding-the-volume-of-a-tetrahedron-by-given-vertices
// For now:
// Sort by input and find the difference squared between the first and last.
const auto inputLess = [](const Coordinate & first, const Coordinate & second)
{ return first.input < second.input; };
// Copy the vector so we don't sort the original.
std::vector<Coordinate> simplexCopy = m_currentSimplex;
std::sort(simplexCopy.begin(), simplexCopy.end(), inputLess);
const auto smallest = simplexCopy.front();
const auto biggest = simplexCopy.back();
return std::pow(biggest.input - smallest.input, 2.0);
}
IterationState SimplexSolver::update()
{
if (m_currentSimplex.size() < 3)
{
throw std::runtime_error("Simplex can't be a line.");
}
// Check for convergence and potentially early return.
// TODO: Make configurable.
const auto volume = calcVolume();
if (volume < 1e-4)
{
return IterationState::CONVERGED;
}
// Do update.
Coordinate potential;
potential.input = newPoint();
potential.cost = m_costFunction.eval(potential.input);
// TODO: Understand why looking at second biggest, not biggest.
// Explanation from paper is that it is in order to make
// the new guess of the next iteration different the current
// biggest.
const auto secondBiggest = *(m_currentSimplex.end() - 2);
if (potential.cost < secondBiggest.cost)
{
// Replace the last simplex value with the better one.
*(m_currentSimplex.end() - 1) = potential;
// TODO: DO I need to sort or is it sorted already?
std::sort(m_currentSimplex.begin(), m_currentSimplex.end());
}
else
{
// Do a contraction.
m_currentSimplex = contract();
}
return IterationState::OK;
}
} // namespace j7s

View File

@ -7,224 +7,33 @@
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#include <cmath>
#include <iostream>
#include <vector>
#include <algorithm>
#include <stdexcept>
#include "j7s-optimization/optimizer.hpp"
class CostFunction
{
public:
CostFunction(double a, double b, double c);
double eval(double input) const;
double actualBest() const;
private:
double m_a;
double m_b;
double m_c;
};
CostFunction::CostFunction(double a, double b, double c) : m_a{a}, m_b{b}, m_c{c} {}
double CostFunction::eval(double input) const
{
return m_a * std::pow(input, 2) + m_b * input + m_c;
}
double CostFunction::actualBest() const
{
// y = a*x**2 + b*x + c
// dy/dx = 2*a*x + b
// 0 = 2*a*x + b
// -b/2*a = x
return (-m_b / (2 * m_a));
}
// A coordinate is pair of a cost input and the resulting the cost.
// We save them together to minimize cost function evaluations.
// This isn't the cleanest thing in the world, but it's worth saving
// evaluations.
struct Coordinate
{
double input;
double cost;
Coordinate() : input{0.0}, cost{0.0} {};
Coordinate(double input, double cost) : input{input}, cost{cost} {};
// Sort by cost.
bool operator<(const Coordinate & other) const { return (cost < other.cost); }
};
enum class IterationState {
OK,
CONVERGED
};
class SimplexSolver
{
public:
SimplexSolver(const CostFunction & costFunction, const std::vector<double> initSimplex);
IterationState update();
Coordinate bestCoord() const;
private:
const CostFunction m_costFunction;
std::vector<Coordinate> m_currentSimplex;
// Helper functions.
double newPoint();
std::vector<Coordinate> contract();
double calcVolume();
};
SimplexSolver::SimplexSolver(const CostFunction & costFunction, const std::vector<double> initSimplex):
m_costFunction{costFunction}
{
m_currentSimplex.reserve(initSimplex.size());
for(const auto val : initSimplex)
{
m_currentSimplex.emplace_back(val, m_costFunction.eval(val));
}
std::sort(m_currentSimplex.begin(), m_currentSimplex.end());
}
Coordinate SimplexSolver::bestCoord() const
{
return m_currentSimplex.front();
}
double SimplexSolver::newPoint()
{
if (m_currentSimplex.size() == 0)
{
throw std::runtime_error("Simplex can't be missing.");
}
// Calculate sum.
double biggest = m_currentSimplex.back().input;
// Calculate volume.
double sum = 0.0;
// All but the biggest, (would be adding 0...)
for (unsigned int index = 0; index < m_currentSimplex.size() - 1; index++)
{
const double diff = m_currentSimplex[index].input - biggest;
sum += diff;
}
const double newPoint = sum * (2.0 / m_currentSimplex.size());
return newPoint;
}
std::vector<Coordinate> SimplexSolver::contract()
{
const auto smallest = m_currentSimplex.front();
std::vector<Coordinate> newVector;
newVector.reserve(m_currentSimplex.size());
newVector.emplace_back(smallest);
// TODO: Really check size before I get here...
for(auto it = m_currentSimplex.begin() + 1; it != m_currentSimplex.end(); it++)
{
const auto oldInput = it->input;
const auto newInput = 0.5*(oldInput + smallest.input);
const auto newCost = m_costFunction.eval(newInput);
newVector.emplace_back(newInput, newCost);
}
std::sort(newVector.begin(), newVector.end());
return newVector;
}
double SimplexSolver::calcVolume()
{
// TODO: For reals do something like:
// https://math.stackexchange.com/questions/337197/finding-the-volume-of-a-tetrahedron-by-given-vertices
// For now:
// Sort by input and find the difference squared between the first and last.
const auto inputLess = [](const Coordinate & first, const Coordinate & second)
{ return first.input < second.input; };
// Copy the vector so we don't sort the original.
std::vector<Coordinate> simplexCopy = m_currentSimplex;
std::sort(simplexCopy.begin(), simplexCopy.end(), inputLess);
const auto smallest = simplexCopy.front();
const auto biggest = simplexCopy.back();
return std::pow(biggest.input - smallest.input, 2.0);
}
IterationState SimplexSolver::update()
{
if (m_currentSimplex.size() < 3)
{
throw std::runtime_error("Simplex can't be a line.");
}
// Check for convergence and potentially early return.
// TODO: Make configurable.
const auto volume = calcVolume();
if(volume < 1e-4)
{
return IterationState::CONVERGED;
}
// Do update.
Coordinate potential;
potential.input = newPoint();
potential.cost = m_costFunction.eval(potential.input);
// TODO: Understand why looking at second biggest, not biggest.
// Explanation from paper is that it is in order to make
// the new guess of the next iteration different the current
// biggest.
const auto secondBiggest = *(m_currentSimplex.end() - 2);
if(potential.cost < secondBiggest.cost)
{
// Replace the last simplex value with the better one.
*(m_currentSimplex.end() - 1) = potential;
// TODO: DO I need to sort or is it sorted already?
std::sort(m_currentSimplex.begin(), m_currentSimplex.end());
}
else
{
// Do a contraction.
m_currentSimplex = contract();
}
return IterationState::OK;
}
#include "j7s-optimization/CostFunction.hpp"
#include "j7s-optimization/SimplexSolver.hpp"
#include "j7s-optimization/common.hpp"
int main(int, char **)
{
const CostFunction cost(2.0, 3.0, 4.0);
const j7s::CostFunction cost(2.0, 3.0, 4.0);
const std::vector<double> init_simplex = {-10, 0, 10};
SimplexSolver solver(cost, init_simplex);
j7s::SimplexSolver solver(cost, init_simplex);
IterationState state = IterationState::OK;
j7s::IterationState state = j7s::IterationState::OK;
for (int cnt = 0; cnt < 1000; cnt++)
{
state = solver.update();
if(state == IterationState::CONVERGED)
if (state == j7s::IterationState::CONVERGED)
{
break;
}
}
if(state == IterationState::CONVERGED)
if (state == j7s::IterationState::CONVERGED)
{
const auto best = solver.bestCoord();
std::cout << "Converged! Best Input: " << best.input << " Cost: " << best.cost << std::endl;
std::cout << "Actual Best: " << cost.actualBest() << " Cost: " << cost.eval(cost.actualBest()) << std::endl;
std::cout << "Actual Best: " << cost.actualBest()
<< " Cost: " << cost.eval(cost.actualBest()) << std::endl;
}
else
{

View File

@ -1,23 +0,0 @@
// Copyright 2022 James Pace
// All Rights Reserved.
//
// For a license to this software contact
// James Pace at jpace121@gmail.com.
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#include "j7s-optimization/optimizer.hpp"
namespace j7s
{
Optimizer::Optimizer()
{
}
Optimizer::~Optimizer()
{
}
} // namespace j7s