• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * This module contains the definition of a Levenberg-Marquardt solver for
3  * non-linear least squares problems of the following form:
4  *
5  *   arg min  ||f(X)||  =  arg min  f(X)^T f(X)
6  *      X                     X
7  *
8  * where X is an Nx1 state vector and f(X) is an Mx1 nonlinear measurement error
9  * function of X which we wish to minimize the norm of.
10  *
11  * Levenberg-Marquardt solves the above problem through a damped Gauss-Newton
12  * method where the solver step on each iteration is chosen such that:
13  *       (J' J + uI) * step = - J' f(x)
14  * where J = df(x)/dx is the Jacobian of the error function, u is a damping
15  * factor, and I is the identity matrix.
16  *
17  * The algorithm implemented here follows Algorithm 3.16 outlined in this paper:
18  * Madsen, Kaj, Hans Bruun Nielsen, and Ole Tingleff.
19  * "Methods for non-linear least squares problems." (2004).
20  *
21  * This algorithm uses a variant of the Marquardt method for updating the
22  * damping factor which ensures that changes in the factor have no
23  * discontinuities or fluttering behavior between solver steps.
24  */
25 #ifndef LOCATION_LBS_CONTEXTHUB_NANOAPPS_COMMON_MATH_LEVENBERG_MARQUARDT_H_
26 #define LOCATION_LBS_CONTEXTHUB_NANOAPPS_COMMON_MATH_LEVENBERG_MARQUARDT_H_
27 
28 #include <stddef.h>
29 #include <stdint.h>
30 
31 #ifdef __cplusplus
32 extern "C" {
33 #endif
34 
35 // Function pointer for computing residual, f(X), and Jacobian, J = df(X)/dX
36 // for a given state value, X, and other parameter data needed for computing
37 // these terms, f_data.
38 //
39 // Note if f_data is not needed, it is allowable for f_data to be passed in
40 // as NULL.
41 //
42 // jacobian is also allowed to be passed in as NULL, and in this case only the
43 // residual will be computed and returned.
44 typedef void (*ResidualAndJacobianFunction)(const float *state,
45                                             const void *f_data,
46                                             float *residual, float *jacobian);
47 
48 
49 #define MAX_LM_STATE_DIMENSION (10)
50 #define MAX_LM_MEAS_DIMENSION (50)
51 
52 // Structure containing fixed parameters for a single LM solve.
53 struct LmParams {
54   // maximum number of iterations allowed by the solver.
55   uint32_t max_iterations;
56 
57   // initial scaling factor for setting the damping factor, i.e.:
58   // damping_factor = initial_u_scale * max(J'J).
59   float initial_u_scale;
60 
61   // magnitude of the cost function gradient required for solution, i.e.
62   // max(gradient) = max(J'f(x)) < gradient_threshold.
63   float gradient_threshold;
64 
65   // magnitude of relative error required for solution step, i.e.
66   // ||step|| / ||state|| < relative_step_thresold.
67   float relative_step_threshold;
68 };
69 
70 // Structure containing data needed for a single LM solve.
71 // Defining the data here allows flexibility in how the memory is allocated
72 // for the solver.
73 struct LmData {
74   float step[MAX_LM_STATE_DIMENSION];
75   float residual[MAX_LM_MEAS_DIMENSION];
76   float residual_new[MAX_LM_MEAS_DIMENSION];
77   float gradient[MAX_LM_MEAS_DIMENSION];
78   float hessian[MAX_LM_STATE_DIMENSION * MAX_LM_STATE_DIMENSION];
79   float temp[MAX_LM_STATE_DIMENSION * MAX_LM_MEAS_DIMENSION];
80 };
81 
82 // Enumeration indicating status of the LM solver.
83 enum LmStatus {
84   RUNNING = 0,
85   // Successful solve conditions:
86   RELATIVE_STEP_SUFFICIENTLY_SMALL,  // solver step is sufficiently small.
87   GRADIENT_SUFFICIENTLY_SMALL,  // cost function gradient is below threshold.
88   HIT_MAX_ITERATIONS,
89 
90   // Solver failure modes:
91   CHOLESKY_FAIL,
92   INVALID_DATA_DIMENSIONS,
93 };
94 
95 // Structure for containing variables needed for a Levenberg-Marquardt solver.
96 struct LmSolver {
97   // Solver parameters.
98   struct LmParams params;
99 
100   // Pointer to solver data.
101   struct LmData *data;
102 
103   // Function for computing residual (f(x)) and jacobian (df(x)/dx).
104   ResidualAndJacobianFunction func;
105 
106   // Number of iterations in current solution.
107   uint32_t num_iter;
108 };
109 
110 // Initializes LM solver with provided parameters and error function.
111 void lmSolverInit(struct LmSolver *solver, const struct LmParams *params,
112                   ResidualAndJacobianFunction func);
113 
114 // Sets pointer for temporary data needed for an individual LM solve.
115 // Note, this must be called prior to calling lmSolverSolve().
116 void lmSolverSetData(struct LmSolver *solver, struct LmData *data);
117 
118 /*
119  * Runs the LM solver for a given set of data and error function.
120  *
121  * INPUTS:
122  * solver : pointer to an struct LmSolver structure.
123  * initial_state : initial guess of the estimation state.
124  * f_data : pointer to additional data needed by the error function.
125  * state_dim : dimension of X.
126  * meas_dim : dimension of f(X), defined in the error function.
127  *
128  * OUTPUTS:
129  * LmStatus : enum indicating state of the solver on completion.
130  * est_state : estimated state.
131  */
132 enum LmStatus lmSolverSolve(struct LmSolver *solver, const float *initial_state,
133                             void *f_data, size_t state_dim, size_t meas_dim,
134                             float *state);
135 
136 ////////////////////////// TEST UTILITIES ////////////////////////////////////
137 // This function is exposed here for testing purposes only.
138 /*
139  * Computes the ratio of actual cost function gain over expected cost
140  * function gain for the given solver step.  This ratio is used to adjust
141  * the solver step size for the next solver iteration.
142  *
143  * INPUTS:
144  * residual: f(x) for the current state x.
145  * residual_new: f(x + step) for the new state after the solver step.
146  * step: the solver step.
147  * gradient: gradient of the cost function F(x).
148  * damping_factor: the current damping factor used in the solver.
149  * state_dim: dimension of the state, x.
150  * meas_dim: dimension of f(x).
151  */
152 float computeGainRatio(const float *residual, const float *residual_new,
153                        const float *step, const float *gradient,
154                        float damping_factor, size_t state_dim,
155                        size_t meas_dim);
156 
157 #ifdef __cplusplus
158 }
159 #endif
160 
161 #endif  // LOCATION_LBS_CONTEXTHUB_NANOAPPS_COMMON_MATH_LEVENBERG_MARQUARDT_H_
162