[if possible, follow-up to BFL]

On Tue, Jan 26, 2010 at 8:17 PM, Ben Podgursky <bpodgursky [..] ...> wrote:

> Hi all,

>

> First, I'm sorry if this is the wrong place to post this; I've tried posting

> in the forum but it's flagging my post as spam no matter how I try to word

> it, so I decided to try here...

weird, personally I always use the mailinglist interface instead of the forum.

> I've been looking into BFL to implement a Kalman filter in a project I'm

> working on. BFL seems to be a great project compared to most filtering

> libraries I've seen--well documented, easy to use.

>

> However, I can't find reference to one feature that some other libraries

> have, an Unscented Kalman filter. I would really prefer using a UKF for my

> project because I'm working with equations generated at runtime, and I

> really don't want to implement the symbolic differentiation I'd need for an

> Extended Kalman Filter...

>

> Does BFL have a UKF implemented that I've missed? And if not, does anyone

> have an idea how difficult it would be for me to extend the library to

> implement one?

I don't think it does, for a simple reason. It's been a while since

I've looked at BFL, but the reason that a UKF isn't in BFL yet is that

a UKF didn't yield the best results for the problems we we're tackling

with BFL. I don't recall the details but you might try google and

search for "tine lefebvre unscented" and something like that.

That said, it should be perfectly possible to implement a UKF in BFL

and IIRC, in one of the above papers, Tine showed that the UKF can be

seen as a special case of what we called linear regression kalman

filter (which might be an intermediate step).

best regards,

Klaas

_______________________________________________

I hereby promise not to top-post on the

BFL mailing list

BFL [..] ...

http://lists.mech.kuleuven.be/mailman/listinfo/bfl

## Unscented Kalman Filter?

On Tue, 26 Jan 2010, Klaas Gadeyne wrote:

> [if possible, follow-up to BFL]

>

> On Tue, Jan 26, 2010 at 8:17 PM, Ben Podgursky <bpodgursky [..] ...> wrote:

>> Hi all,

>>

>> First, I'm sorry if this is the wrong place to post this; I've tried posting

>> in the forum but it's flagging my post as spam no matter how I try to word

>> it, so I decided to try here...

>

> weird, personally I always use the mailinglist interface instead of the forum.

>

>> I've been looking into BFL to implement a Kalman filter in a project I'm

>> working on. BFL seems to be a great project compared to most filtering

>> libraries I've seen--well documented, easy to use.

>>

>> However, I can't find reference to one feature that some other libraries

>> have, an Unscented Kalman filter. I would really prefer using a UKF for my

>> project because I'm working with equations generated at runtime, and I

>> really don't want to implement the symbolic differentiation I'd need for an

>> Extended Kalman Filter...

>>

>> Does BFL have a UKF implemented that I've missed? And if not, does anyone

>> have an idea how difficult it would be for me to extend the library to

>> implement one?

>

> I don't think it does, for a simple reason. It's been a while since

> I've looked at BFL, but the reason that a UKF isn't in BFL yet is that

> a UKF didn't yield the best results for the problems we we're tackling

> with BFL. I don't recall the details but you might try google and

> search for "tine lefebvre unscented" and something like that.

>

> That said, it should be perfectly possible to implement a UKF in BFL

> and IIRC, in one of the above papers, Tine showed that the UKF can be

> seen as a special case of what we called linear regression kalman

> filter (which might be an intermediate step).

UKF can also be considered to be a particle filter with a very small, but

selected set of samples. The PF (re)sampling functionalities could be

reused, I guess. If you decide to make a contribution to BFL, I would

suggest to make it under a somewhat more general form, that of

pseudo-random filters; UKF is a special case of those.

Herman

## Unscented Kalman Filter?

On Tue, 26 Jan 2010, Klaas Gadeyne wrote:

> [if possible, follow-up to BFL]

>

> On Tue, Jan 26, 2010 at 8:17 PM, Ben Podgursky <bpodgursky [..] ...> wrote:

>> Hi all,

>>

>> First, I'm sorry if this is the wrong place to post this; I've tried posting

>> in the forum but it's flagging my post as spam no matter how I try to word

>> it, so I decided to try here...

>

> weird, personally I always use the mailinglist interface instead of the forum.

>

>> I've been looking into BFL to implement a Kalman filter in a project I'm

>> working on. BFL seems to be a great project compared to most filtering

>> libraries I've seen--well documented, easy to use.

>>

>> However, I can't find reference to one feature that some other libraries

>> have, an Unscented Kalman filter. I would really prefer using a UKF for my

>> project because I'm working with equations generated at runtime, and I

>> really don't want to implement the symbolic differentiation I'd need for an

>> Extended Kalman Filter...

>>

>> Does BFL have a UKF implemented that I've missed? And if not, does anyone

>> have an idea how difficult it would be for me to extend the library to

>> implement one?

>

> I don't think it does, for a simple reason. It's been a while since

> I've looked at BFL, but the reason that a UKF isn't in BFL yet is that

> a UKF didn't yield the best results for the problems we we're tackling

> with BFL. I don't recall the details but you might try google and

> search for "tine lefebvre unscented" and something like that.

>

> That said, it should be perfectly possible to implement a UKF in BFL

> and IIRC, in one of the above papers, Tine showed that the UKF can be

> seen as a special case of what we called linear regression kalman

> filter (which might be an intermediate step).

UKF can also be considered to be a particle filter with a very small, but

selected set of samples. The PF (re)sampling functionalities could be

reused, I guess. If you decide to make a contribution to BFL, I would

suggest to make it under a somewhat more general form, that of

pseudo-random filters; UKF is a special case of those.

Herman

## Unscented Kalman Filter?

[dropped orocos-dev from cc:]

On Wed, Jan 27, 2010 at 7:58 AM, Herman Bruyninckx

<Herman [dot] Bruyninckx [..] ...> wrote:

[...]

> UKF can also be considered to be a particle filter with a very small, but

> selected set of samples. The PF (re)sampling functionalities could be

> reused, I guess.

I don't think so, because the (sampling algorithms from) PFs in BFL

are used for truly random filters (yes that's a matter of definition,

but it is the one I used :-).

Side note: note however that an unscented particle filter also exists

which is a true particle filter but still uses the unscented transform

<quote: bluntly copied from "somewhere on the internet", hope the

copyright owner KULeuven doesn't sue me for copyright violation :-)

Other variations found in literature are e.g. the Sigma Point

Particle Filter

and the Gaussian Mixture Sigma Point Particle Filter (van der Merwe and

Wan 2003). These are Particle Filters that use a Linear Regression Kalman

Filter (Lefebvre 2003) in their proposal step. Worth mentioning in this scope

is the Unscented Particle Filter (van der Merwe, Doucet, de Freytas, and

Wan 2000; Julier and Uhlmann 2004): The Unscented Kalman Filter (UKF)

(Julier and Uhlmann 1997) is known to always generate consistent estimates,

at the cost of a sometimes rather uninformative state estimate (Lefebvre,

Bruyninckx, and De Schutter 2004b). This makes it extremely valuable

since it generates proposal densities with heavier tails than the true posterior

(See Section 3.5.4). This reduces the Monte Carlo variance of the resulting

estimator, or allows to obtain the same variance with less samples

(i.e. faster).

As will be demonstrated in Chapter 6, the large computational complexity

is still a problem for the ACM tasks dealt with in this thesis,

especially if larger

uncertainties are considered. Therefore, the use of the above described filters

for estimation in ACM is a topic of future research.

> If you decide to make a contribution to BFL, I would

> suggest to make it under a somewhat more general form, that of

> pseudo-random filters;

As Peter says: _Any_ contribution is useful.

IIRC a PhD student at the CS department already implemented a class of

pseudo random filters for BFL, but they were never contributed...

Klaas

_______________________________________________

I hereby promise not to top-post on the

BFL mailing list

BFL [..] ...

http://lists.mech.kuleuven.be/mailman/listinfo/bfl

## Unscented Kalman Filter?

Hi all,

First, I'm sorry if this is the wrong place to post this; I've tried posting

in the forum but it's flagging my post as spam no matter how I try to word

it, so I decided to try here...

I've been looking into BFL to implement a Kalman filter in a project I'm

working on. BFL seems to be a great project compared to most filtering

libraries I've seen--well documented, easy to use.

However, I can't find reference to one feature that some other libraries

have, an Unscented Kalman filter. I would really prefer using a UKF for my

project because I'm working with equations generated at runtime, and I

really don't want to implement the symbolic differentiation I'd need for an

Extended Kalman Filter...

Does BFL have a UKF implemented that I've missed? And if not, does anyone

have an idea how difficult it would be for me to extend the library to

implement one?

Thanks in advance,

Ben

## Unscented Kalman Filter?

[if possible, follow-up to BFL]

On Tue, Jan 26, 2010 at 8:17 PM, Ben Podgursky <bpodgursky [..] ...> wrote:

> Hi all,

>

> First, I'm sorry if this is the wrong place to post this; I've tried posting

> in the forum but it's flagging my post as spam no matter how I try to word

> it, so I decided to try here...

weird, personally I always use the mailinglist interface instead of the forum.

> I've been looking into BFL to implement a Kalman filter in a project I'm

> working on. BFL seems to be a great project compared to most filtering

> libraries I've seen--well documented, easy to use.

>

> However, I can't find reference to one feature that some other libraries

> have, an Unscented Kalman filter. I would really prefer using a UKF for my

> project because I'm working with equations generated at runtime, and I

> really don't want to implement the symbolic differentiation I'd need for an

> Extended Kalman Filter...

>

> Does BFL have a UKF implemented that I've missed? And if not, does anyone

> have an idea how difficult it would be for me to extend the library to

> implement one?

I don't think it does, for a simple reason. It's been a while since

I've looked at BFL, but the reason that a UKF isn't in BFL yet is that

a UKF didn't yield the best results for the problems we we're tackling

with BFL. I don't recall the details but you might try google and

search for "tine lefebvre unscented" and something like that.

That said, it should be perfectly possible to implement a UKF in BFL

and IIRC, in one of the above papers, Tine showed that the UKF can be

seen as a special case of what we called linear regression kalman

filter (which might be an intermediate step).

best regards,

Klaas

## Unscented Kalman Filter?

Hi, thanks!

I think I've found the paper you mentioned: "Kalman Filters for nonlinear

systems:

a comparison of performance."... let me know if I'm wrong about that.

I'll take a look at that paper and the library design. If I'm feeling

ambitious, I might take a shot at implementing a UKF...

Thanks,

Ben

On Tue, Jan 26, 2010 at 1:48 PM, Klaas Gadeyne <klaas [dot] gadeyne [..] ...>wrote:

> [if possible, follow-up to BFL]

>

> On Tue, Jan 26, 2010 at 8:17 PM, Ben Podgursky <bpodgursky [..] ...>

> wrote:

> > Hi all,

> >

> > First, I'm sorry if this is the wrong place to post this; I've tried

> posting

> > in the forum but it's flagging my post as spam no matter how I try to

> word

> > it, so I decided to try here...

>

> weird, personally I always use the mailinglist interface instead of the

> forum.

>

> > I've been looking into BFL to implement a Kalman filter in a project I'm

> > working on. BFL seems to be a great project compared to most filtering

> > libraries I've seen--well documented, easy to use.

> >

> > However, I can't find reference to one feature that some other libraries

> > have, an Unscented Kalman filter. I would really prefer using a UKF for

> my

> > project because I'm working with equations generated at runtime, and I

> > really don't want to implement the symbolic differentiation I'd need for

> an

> > Extended Kalman Filter...

> >

> > Does BFL have a UKF implemented that I've missed? And if not, does

> anyone

> > have an idea how difficult it would be for me to extend the library to

> > implement one?

>

> I don't think it does, for a simple reason. It's been a while since

> I've looked at BFL, but the reason that a UKF isn't in BFL yet is that

> a UKF didn't yield the best results for the problems we we're tackling

> with BFL. I don't recall the details but you might try google and

> search for "tine lefebvre unscented" and something like that.

>

> That said, it should be perfectly possible to implement a UKF in BFL

> and IIRC, in one of the above papers, Tine showed that the UKF can be

> seen as a special case of what we called linear regression kalman

> filter (which might be an intermediate step).

>

> best regards,

>

> Klaas

>

_______________________________________________

I hereby promise not to top-post on the

BFL mailing list

BFL [..] ...

http://lists.mech.kuleuven.be/mailman/listinfo/bfl

## description change