Skip to content

D
dlib
Commit c2196598 authored by Davis King's avatar Davis King
Browse files
Options

Filled out spec for rls_filter and added asserts

parent f3b0e8bc
Hide whitespace changes
Inline Side-by-side
Showing with 97 additions and 0 deletions
dlib/filtering/rls_filter.h
View file @ c2196598
......@@ -43,6 +43,17 @@ namespace dlib
double C = 100
)
{
// make sure requires clause is not broken
DLIB_ASSERT(0 < forget_factor && forget_factor <= 1 &&
0 < C && size_ >= 2,
"\t rls_filter::rls_filter()"
<< "\n\t invalid arguments were given to this function"
<< "\n\t forget_factor: " << forget_factor
<< "\n\t C: " << C
<< "\n\t size_: " << size_
<< "\n\t this: " << this
);
size = size_;
count = 0;
filter = rls(forget_factor, C);
......@@ -89,6 +100,18 @@ namespace dlib
const matrix_exp<EXP>& z
)
{
// make sure requires clause is not broken
DLIB_ASSERT(is_col_vector(z) == true &&
z.size() != 0 &&
(get_predicted_next_state().size()==0 || z.size()==get_predicted_next_state().size()),
"\t void rls_filter::update(z)"
<< "\n\t invalid arguments were given to this function"
<< "\n\t is_col_vector(z): " << is_col_vector(z)
<< "\n\t z.size(): " << z.size()
<< "\n\t get_predicted_next_state().size(): " << get_predicted_next_state().size()
<< "\n\t this: " << this
);
// initialize data if necessary
if (data.size() == 0)
{
......
dlib/filtering/rls_filter_abstract.h
View file @ c2196598
......@@ -15,6 +15,20 @@ namespace dlib
{
/*!
WHAT THIS OBJECT REPRESENTS
This object is a tool for doing time series prediction using linear
recursive least squares. In particular, this object takes a sequence
of points from the user and, at each step, attempts to predict the
value of the next point.
To accomplish this, this object maintains a fixed size buffer of recent
points. Each prediction is a linear combination of the points in this
history buffer. It uses the recursive least squares algorithm to
determine how to best combine the contents of the history buffer to
predict each point. Therefore, each time update() is called with
a point, recursive least squares updates the linear combination weights,
and then we insert the point into the history buffer. After that, the
next prediction is based on these updated weights and the current history
buffer.
!*/
public:
......@@ -26,6 +40,7 @@ namespace dlib
- #get_window_size() == 5
- #get_c() == 100
- #get_forget_factor() == 0.8
- #get_predicted_next_state().size() == 0
!*/
explicit rls_filter (
......@@ -42,29 +57,88 @@ namespace dlib
- #get_window_size() == size
- #get_forget_factor() == forget_factor
- #get_c() == C
- #get_predicted_next_state().size() == 0
!*/
double get_c(
) const;
/*!
ensures
- returns the regularization parameter. It is the parameter that determines
the trade-off between trying to fit the data points given to update() or
allowing more errors but hopefully improving the generalization of the
predictions. Larger values encourage exact fitting while smaller values
of C may encourage better generalization.
!*/
double get_forget_factor(
) const;
/*!
ensures
- This object uses exponential forgetting in its implementation of recursive
least squares. Therefore, this function returns the "forget factor".
- if (get_forget_factor() == 1) then
- In this case, exponential forgetting is disabled.
- The recursive least squares algorithm will implicitly take all previous
calls to update(z) into account when estimating the optimal weights for
linearly combining the history buffer into a prediction of the next point.
- else
- Old calls to update(z) are eventually forgotten. That is, the smaller
the forget factor, the less the recursive least squares algorithm will
care about attempting to find linear combination weights which would have
make good predictions on old points. It will care more about fitting
recent points. This is appropriate if the statistical properties of
the time series we are modeling are not constant.
!*/
unsigned long get_window_size (
) const;
/*!
ensures
- returns the size of the history buffer. This is the number of points which
are linearly combine to make the predictions returned by get_predicted_next_state().
!*/
void update (
);
/*!
ensures
- Propagates the prediction forward in time.
- In particular, the value in #get_predicted_next_state() is inserted
into the history buffer and then the next prediction is estimated
based on this updated history buffer.
- #get_predicted_next_state() == the prediction for the next point
in the time series.
!*/
template <typename EXP>
void update (
const matrix_exp<EXP>& z
);
/*!
requires
- is_col_vector(z) == true
- z.size() != 0
- if (get_predicted_next_state().size() != 0) then
- z.size() == get_predicted_next_state().size()
(i.e. z must be the same size as all the previous z values given
to this function)
ensures
- Updates the state of this filter based on the current measurement in z.
- In particular, the filter weights are updated and z is inserted into
the history buffer. Then the next prediction is estimated based on
these updated weights and history buffer.
- #get_predicted_next_state() == the prediction for the next point
in the time series.
!*/
const matrix<double,0,1>& get_predicted_next_state(
);
/*!
ensures
- returns the estimate of the next point we will observe in the
time series data.
!*/
};
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment