File size: 3,344 Bytes
fe2b563
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
#ifndef ANDERSONACCELERATION_H_
#define ANDERSONACCELERATION_H_

#include "Types.h"
#include <cassert>
#include <algorithm>
#include <vector>
#include <omp.h>
#include <fstream>

class AndersonAcceleration
{
public:
	AndersonAcceleration()
		:m_(-1), dim_(-1), iter_(-1), col_idx_(-1) {}

	void replace(const Scalar *u)
	{
		current_u_ = Eigen::Map<const VectorX>(u, dim_);
	}

	const VectorX& compute(const Scalar* g)
	{
		assert(iter_ >= 0);

		Eigen::Map<const VectorX> G(g, dim_);
		current_F_ = G - current_u_;

		if (iter_ == 0)
		{
			prev_dF_.col(0) = -current_F_;
			prev_dG_.col(0) = -G;
			current_u_ = G;
		}
		else
		{
			prev_dF_.col(col_idx_) += current_F_;
			prev_dG_.col(col_idx_) += G;

			Scalar eps = 1e-14;
			Scalar scale = std::max(eps, prev_dF_.col(col_idx_).norm());
			dF_scale_(col_idx_) = scale;
			prev_dF_.col(col_idx_) /= scale;

			int m_k = std::min(m_, iter_);


			if (m_k == 1)
			{
				theta_(0) = 0;
				Scalar dF_sqrnorm = prev_dF_.col(col_idx_).squaredNorm();
				M_(0, 0) = dF_sqrnorm;
				Scalar dF_norm = std::sqrt(dF_sqrnorm);

                if (dF_norm > eps) {
					theta_(0) = (prev_dF_.col(col_idx_) / dF_norm).dot(current_F_ / dF_norm);
				}
			}
			else
			{
				// Update the normal equation matrix, for the column and row corresponding to the new dF column
				VectorX new_inner_prod = (prev_dF_.col(col_idx_).transpose() * prev_dF_.block(0, 0, dim_, m_k)).transpose();
				M_.block(col_idx_, 0, 1, m_k) = new_inner_prod.transpose();
				M_.block(0, col_idx_, m_k, 1) = new_inner_prod;

				// Solve normal equation
				cod_.compute(M_.block(0, 0, m_k, m_k));
				theta_.head(m_k) = cod_.solve(prev_dF_.block(0, 0, dim_, m_k).transpose() * current_F_);
			}

			// Use rescaled theata to compute new u
			current_u_ = G - prev_dG_.block(0, 0, dim_, m_k) * ((theta_.head(m_k).array() / dF_scale_.head(m_k).array()).matrix());
			col_idx_ = (col_idx_ + 1) % m_;
			prev_dF_.col(col_idx_) = -current_F_;
			prev_dG_.col(col_idx_) = -G;
		}

		iter_++;
		return current_u_;
	}
    void reset(const Scalar *u)
    {
        iter_ = 0;
        col_idx_ = 0;
        current_u_ = Eigen::Map<const VectorX>(u, dim_);
    }

	// m: number of previous iterations used
	// d: dimension of variables
	// u0: initial variable values
	void init(int m, int d, const Scalar* u0)
	{
		assert(m > 0);
		m_ = m;
		dim_ = d;
		current_u_.resize(d);
		current_F_.resize(d);
		prev_dG_.resize(d, m);
		prev_dF_.resize(d, m);
		M_.resize(m, m);
		theta_.resize(m);
		dF_scale_.resize(m);
		current_u_ = Eigen::Map<const VectorX>(u0, d);
		iter_ = 0;
		col_idx_ = 0;
	}

private:
	VectorX current_u_;
	VectorX current_F_;
	MatrixXX prev_dG_;
	MatrixXX prev_dF_;
	MatrixXX M_;		// Normal equations matrix for the computing theta
	VectorX	theta_;	// theta value computed from normal equations
	VectorX dF_scale_;		// The scaling factor for each column of prev_dF
	Eigen::CompleteOrthogonalDecomposition<MatrixXX> cod_;

	int m_;		// Number of previous iterates used for Andreson Acceleration
	int dim_;	// Dimension of variables
	int iter_;	// Iteration count since initialization
	int col_idx_;	// Index for history matrix column to store the next value
	int m_k_;

	Eigen::Matrix4d current_T_;
	Eigen::Matrix4d current_F_T_;

	MatrixXX T_prev_dF_;
	MatrixXX T_prev_dG_;
};


#endif /* ANDERSONACCELERATION_H_ */