xref: /petsc/src/tao/linesearch/impls/armijo/armijo.h (revision bebe2cf65d55febe21a5af8db2bd2e168caaa2e7)
1 #ifndef __TAOLINESEARCH_ARMIJO_H
2 #define __TAOLINESEARCH_ARMIJO_H
3 
4 /* Context for an Armijo (nonmonotone) linesearch for unconstrained
5    minimization.
6 
7    Given a function f, the current iterate x, and a descent direction d:
8    Find the smallest i in 0, 1, 2, ..., such that:
9 
10       f(x + (beta**i)d) <= f(x) + (sigma*beta**i)<grad f(x),d>
11 
12    The nonmonotone modification of this linesearch replaces the f(x) term
13    with a reference value, R, and seeks to find the smallest i such that:
14 
15       f(x + (beta**i)d) <= R + (sigma*beta**i)<grad f(x),d>
16 
17    This modification does effect neither the convergence nor rate of
18    convergence of an algorithm when R is chosen appropriately.  Essentially,
19    R must decrease on average in some sense.  The benefit of a nonmonotone
20    linesearch is that local minimizers can be avoided (by allowing increase
21    in function value), and typically, fewer iterations are performed in
22    the main code.
23 
24    The reference value is chosen based upon some historical information
25    consisting of function values for previous iterates.  The amount of
26    historical information used is determined by the memory size where the
27    memory is used to store the previous function values.  The memory is
28    initialized to alpha*f(x^0) for some alpha >= 1, with alpha=1 signifying
29    that we always force decrease from the initial point.
30 
31    The reference value can be the maximum value in the memory or can be
32    chosen to provide some mean descent.  Elements are removed from the
33    memory with a replacement policy that either removes the oldest
34    value in the memory (FIFO), or the largest value in the memory (MRU).
35 
36    Additionally, we can add a watchdog strategy to the search, which
37    essentially accepts small directions and only checks the nonmonotonic
38    descent criteria every m-steps.  This strategy is NOT implemented in
39    the code.
40 
41    Finally, care must be taken when steepest descent directions are used.
42    For example, when the Newton direction is not not satisfy a sufficient
43    descent criteria.  The code will apply the same test regardless of
44    the direction.  This type of search may not be appropriate for all
45    algorithms.  For example, when a gradient direction is used, we may
46    want to revert to the best point found and reset the memory so that
47    we stay in an appropriate level set after using a gradient steps.
48    This type of search is currently NOT supported by the code.
49 
50    References:
51     Armijo, "Minimization of Functions Having Lipschitz Continuous
52       First-Partial Derivatives," Pacific Journal of Mathematics, volume 16,
53       pages 1-3, 1966.
54     Ferris and Lucidi, "Nonmonotone Stabilization Methods for Nonlinear
55       Equations," Journal of Optimization Theory and Applications, volume 81,
56       pages 53-71, 1994.
57     Grippo, Lampariello, and Lucidi, "A Nonmonotone Line Search Technique
58       for Newton's Method," SIAM Journal on Numerical Analysis, volume 23,
59       pages 707-716, 1986.
60     Grippo, Lampariello, and Lucidi, "A Class of Nonmonotone Stabilization
61       Methods in Unconstrained Optimization," Numerische Mathematik, volume 59,
62       pages 779-805, 1991. */
63 #include <petsc/private/taolinesearchimpl.h>
64 typedef struct {
65   PetscReal *memory;
66 
67   PetscReal alpha;                      /* Initial reference factor >= 1 */
68   PetscReal beta;                       /* Steplength determination < 1 */
69   PetscReal beta_inf;           /* Steplength determination < 1 */
70   PetscReal sigma;                      /* Acceptance criteria < 1) */
71   PetscReal minimumStep;                /* Minimum step size */
72   PetscReal lastReference;              /* Reference value of last iteration */
73 
74   PetscInt memorySize;          /* Number of functions kept in memory */
75   PetscInt current;                     /* Current element for FIFO */
76   PetscInt referencePolicy;             /* Integer for reference calculation rule */
77   PetscInt replacementPolicy;   /* Policy for replacing values in memory */
78 
79   PetscBool nondescending;
80   PetscBool memorySetup;
81 
82   Vec x;        /* Maintain reference to variable vector to check for changes */
83   Vec work;
84 } TaoLineSearch_ARMIJO;
85 
86 #endif
87