Help us with this petition for security on anti-collision systems

On Tuesday, May 26, 2015 at 7:44:07 PM UTC-4, Lucas wrote:

Tango Eight, your statement is lacking of a scientific base: WHAT demonstrates that the "prediction algorithm works very well" ?

*Extensive* end user experience.

This might be helpful: http://en.wikipedia.org/wiki/Straw_man

best regards,
Evan Ludeman / T8

What Flarm calls "prediction" I think that most likely is a simple
projection. It is quite likely calculated worst than how we calculate the
best point to turn in thermal.
I am referring to the "Beep" in Zanders, or in some flight computers .
If you want to see how a "prediction" is working, look at the thermal
Orbiter I have programmed
https://github.com/LK8000/LK8000/blo...lc/Orbiter.cpp
which is quite similar to what Zander and SeeYou Mobile (and possibly other
software, I don't really know) do.
This is a prediction based on turning angle, estimated banking etc. and I
mention it here for a reason:
there is floating point math involved in such kind of predictions.
We use 400mhz or best ARM cpu on PNA-PDAs.
Flarm is tuned to "predict" on a 8mhz CPU by Atmel, a reduced instruction
set microcontroller that has no math coprocessor and cannot do floating
point calculations natively.

A prediction seems like something magic, and I doubt this is the case.
Each device (flarm, dsx) transmits its own position "predicted" with a
simple projection for the next second .
If your own device matches its own "predicted" position with the one
received from another one, it beeps.
That's how it works.
A projection cannot predict when you level and go straight, nevertheless as
you say it works .
It can not work "very well", as you say. But it is better than nothing.

The assumption is that the glider in thermal with you, or arriving in front
of you, has a device with the same protocol.
In the alps this is no more granted. This is what this thread is about.

greets
paolo




"Tango Eight" wrote in message
...

On Tuesday, May 26, 2015 at 7:44:07 PM UTC-4, Lucas wrote:

Tango Eight, your statement is lacking of a scientific base: WHAT
demonstrates that the "prediction algorithm works very well" ?

*Extensive* end user experience.

This might be helpful: http://en.wikipedia.org/wiki/Straw_man

best regards,
Evan Ludeman / T8

On Wednesday, May 27, 2015 at 7:46:25 PM UTC-7, pcool wrote:

The assumption is that the glider in thermal with you, or arriving in front
of you, has a device with the same protocol.
In the alps this is no more granted. This is what this thread is about.

Wow, why would people buy an incompatible device when there are multiple manufacturers of compatible devices in the market?

On Thursday, 28 May 2015 07:05:42 UTC+2, Andy Blackburn wrote:
Wow, why would people buy an incompatible device when there are multiple manufacturers of compatible devices in the market?

What would happen if the relationship soured between FLARM and the manufacturer of your chosen FLARM device?
Flarm could easily issue another upgrade to the protocol and you are left with a $1000+ system which is now totally worthless and useless.

Who is incompatible with who? You have the freedom to choose a device
manufacturer.
The TAdvisor, and probably the OGN devices soon, are not worst than flarm to
do this job.
Anyway, as a wise guy ("Buddy Bob") here stated, shortly we may have OGN
devices acting as collision avoidance systems.
At that point Flarm will change its protocol and adopt the open one.

I fully agree with Bob, it is pointless to ask Flarm to open the protocol.
What we need is several other manufacturers selling their own devices, based
on the OGN open software for example. I have not signed the petition for
this reason.


"Andy Blackburn" wrote in message
...

On Wednesday, May 27, 2015 at 7:46:25 PM UTC-7, pcool wrote:

The assumption is that the glider in thermal with you, or arriving in
front
of you, has a device with the same protocol.
In the alps this is no more granted. This is what this thread is about.

Wow, why would people buy an incompatible device when there are multiple
manufacturers of compatible devices in the market?

On Tuesday, May 26, 2015 at 4:44:07 PM UTC-7, Lucas wrote:
The T-Advisor AND Flarm are NOT anticollision system.
Anticollision systems are those who tell the pilot what to do (like TCAS), in case of emergency. Neither system does this.
BOTH systems are NOT anticollision systems.

This is a bit of a hair-splitting argument, but to be clear - there a 1) Traffic display systems (show traffic within a detection volume, but provide no alerts), 2) Traffic advisory systems (alert to new traffic entering a detection volume - like PCAS), 3) Collision detection systems (calculate and warn of other aircraft on a probable collision path - like FLARM), 4) Anti-collision systems (advise pilots on action to avoid a collision), 5) Automated anti-collision systems (autonomously take action to avoid collisions - I'm not aware of any of these - outside of military terrain-following autopilots).

Usefulness goes up as you move up the hierarchy. IMO FLARM, being higher that other systems is more useful.


Cleared this, those who talk about the "predictive algorithm", can please explain:

1) how it works, since they must know how it works, to be in the position of saying that it works or it doesn't

2) how they know that such an algorithm has been implemented into a Flarm system: what proof do they have of this ?

Really? That's a serious question? Well, it warns me of converging traffic and when I look, there is in fact converging traffic in the direction indicated. There is no traffic converging on me for which I get no warning and warnings for traffic that is not a series factor is almost nonexistent.


3) if they have ever seen the trace of at least 10 glider flights in different conditions (competition, cross country, around-the-airfield, ridge soaring,....); whoever has seen some, not many, traces of flights, without the need to be a glider pilot, can understand that a prediction of the position of a glider in a future time beyond a fistful of seconds is impossible, exactly IMPOSSIBLE, since not even its pilot knows it, apart from some cases, like straight flying and constant turn rate thermaling. A glider pilot knows that he will be changing the trajectory of the glider to search for the best netto value, which depends on the micro air movements, which are unknown to the pilot in terms of exact location. Is there a machine capable of predicting these locations ? And even if there was, is there a machine capable of predicting what a glider pilot will do in the next 30-60-80 seconds ? Because this is what the rumored (never verified) "prediction algorithm" does. This is spectacular indeed !

Impossible? As a control-systems engineer I can tell you for a fact that a 1 second sample rate is perfectly adequate for this purpose and you only need 2-3 good data points for each aircraft to make a decent prediction. Even with dropped packets this is a reasonable task. Glider flight dynamics are not so abrupt as to make this an impossible task and pilots are not generally making so many aggressive control inputs as to flail the system. FLARM uses a probabilistic approach base on total energy to err on the side of possible control inputs that handles most situations well.


4) even if they found a system to predict the position of the glider with a certain probability, would they trust as optimal a system that has (obviously) a probability to fail the prediction and miss a danger of collision ? Even if the probability was low (all but sure, since never demonstrated with objective tests and calculated data), 2, 3, 10 collisions (and deads) out of XX'XXX flights are too much. In aeronautics, this approach is wrong: this is not the way we work in professional aeronautics, that has taken us where we are in aviation

We don't need optimal, we need better than human perception and FLARM does that very well indeed.


9B

On Wednesday, May 27, 2015 at 10:35:46 PM UTC+2, Andy Blackburn wrote:
On Tuesday, May 26, 2015 at 4:44:07 PM UTC-7, Lucas wrote:
The T-Advisor AND Flarm are NOT anticollision system.
Anticollision systems are those who tell the pilot what to do (like TCAS), in case of emergency. Neither system does this.
BOTH systems are NOT anticollision systems.

This is a bit of a hair-splitting argument, but to be clear - there a 1) Traffic display systems (show traffic within a detection volume, but provide no alerts), 2) Traffic advisory systems (alert to new traffic entering a detection volume - like PCAS), 3) Collision detection systems (calculate and warn of other aircraft on a probable collision path - like FLARM), 4) Anti-collision systems (advise pilots on action to avoid a collision), 5) Automated anti-collision systems (autonomously take action to avoid collisions - I'm not aware of any of these - outside of military terrain-following autopilots).

Usefulness goes up as you move up the hierarchy. IMO FLARM, being higher that other systems is more useful.


Cleared this, those who talk about the "predictive algorithm", can please explain:

1) how it works, since they must know how it works, to be in the position of saying that it works or it doesn't

2) how they know that such an algorithm has been implemented into a Flarm system: what proof do they have of this ?

Really? That's a serious question? Well, it warns me of converging traffic and when I look, there is in fact converging traffic in the direction indicated. There is no traffic converging on me for which I get no warning and warnings for traffic that is not a series factor is almost nonexistent.


3) if they have ever seen the trace of at least 10 glider flights in different conditions (competition, cross country, around-the-airfield, ridge soaring,....); whoever has seen some, not many, traces of flights, without the need to be a glider pilot, can understand that a prediction of the position of a glider in a future time beyond a fistful of seconds is impossible, exactly IMPOSSIBLE, since not even its pilot knows it, apart from some cases, like straight flying and constant turn rate thermaling. A glider pilot knows that he will be changing the trajectory of the glider to search for the best netto value, which depends on the micro air movements, which are unknown to the pilot in terms of exact location. Is there a machine capable of predicting these locations ? And even if there was, is there a machine capable of predicting what a glider pilot will do in the next 30-60-80 seconds ? Because this is what the rumored (never verified) "prediction algorithm" does. This is spectacular indeed !

Impossible? As a control-systems engineer I can tell you for a fact that a 1 second sample rate is perfectly adequate for this purpose and you only need 2-3 good data points for each aircraft to make a decent prediction. Even with dropped packets this is a reasonable task. Glider flight dynamics are not so abrupt as to make this an impossible task and pilots are not generally making so many aggressive control inputs as to flail the system. FLARM uses a probabilistic approach base on total energy to err on the side of possible control inputs that handles most situations well.


4) even if they found a system to predict the position of the glider with a certain probability, would they trust as optimal a system that has (obviously) a probability to fail the prediction and miss a danger of collision ? Even if the probability was low (all but sure, since never demonstrated with objective tests and calculated data), 2, 3, 10 collisions (and deads) out of XX'XXX flights are too much. In aeronautics, this approach is wrong: this is not the way we work in professional aeronautics, that has taken us where we are in aviation

We don't need optimal, we need better than human perception and FLARM does that very well indeed.


9B

Andy, the T-Advisor is a
3) Collision detection system
Obviously there are different ways to detect a possible collision: to be "extreme", just to make the difference evident illustrating the two extreme positions, you can predict a collision based on approach speeds (very objective and 100% sure data) or using a crystal ball (0% objective and totally unsure data). There are ways in between, obviously.
If you possibly had the responsibility of a life, I am sure you would not go for a "probabilistic" approach, but you would base the functioning of the system only on 100% sure data.
And for a collision, the only sure data are the relative approach speeds and distance, hence time to impact. The rest are speculations about what the pilot will or will not do: you can be in a trajectory considered safe for a predictive system and suddently feel the right wing raising and deciding to turn right, where, by chance, there is another plane in your blind spot. No system can predict this.

When you write:
"it warns me of converging traffic and when I look, there is in fact converging traffic in the direction indicated. There is no traffic converging on me for which I get no warning and warnings for traffic that is not a series factor is almost nonexistent."
That is exactly the same output you get from a T-Advisor, based on the approach speed and distance of the surrounding planes.
This should show you that what you wrote doesn't demonstrate that any "prediction algorithm" is in place.
So you don't know what is implemented in Flarm because, on the contrary of the systems developed with an aeronautical procedure, it's not verified by anyone.

"Prediction": you are a control system engineer, therefore you work with deterministic systems. The glider pilot, whose actions depend on atmospheric conditions unknown a priori (i.e. vertical gusts/thermals), is not a deterministic system.
I agree with you that it is possible to say with an approximate precision where a glider will be in 2-3-4 seconds, because it's dynamic is slow compared to this time length and, in few seconds, the most abrupt control inputs from the pilot will not deviate it a lot from the existing trajectory. But it is clear to any glider pilot or person who sees a flight trace that affirming that a system can "predict" where the glider will be in 20-40 seconds anytime, is like saying that it has a real crystal ball.
It looks like discussing of the obvious: take a flight trace and place yourself in any position along it and then say if you could have told where the glider would be after 20-40 seconds. Leaving aside the straight glides and regular thermals.
This is indisputable, I think.

We don't need optimal, we need better than human perception and FLARM does that very well indeed.
If you had to design a system to save the life of two persons, I am sure you would not take a probabilistic approach, but you would take a path based on 100% sure data. Since these are available (approach speed/distance = time to impact, although calibrated with some other parameters like relative altitude and others).
All system are improving human perception: this is not the point.
The point is which approach is safer, surer for avoiding a collision. An engineering discussion.

On Friday, May 29, 2015 at 2:13:45 AM UTC-7, Lucas wrote:
you are a control system engineer, therefore you work with deterministic systems. The glider pilot, whose actions depend on atmospheric conditions unknown a priori (i.e. vertical gusts/thermals), is not a deterministic system..

Nope - control systems engineers spend most of their time dealing with the effects of noise and non-determinist effects - no system is perfectly deterministic and you assume so at your peril. The issue is to figure out "how" deterministic a system is . Also, if you make NO assumptions about aircraft dynamics many systems are not accurately observable or controllable. Knowing what the aircraft can do dynamically makes a huge difference in knowing where it is and where it can possibly go. Assuming nothing about aircraft dynamics in your system IS making an assumption - you assume the system will keep doing exactly what it was doing when you measured it last - if it was going up it will keep going up indefinitely, if it was circling you assuming it will follow the instantaneous velocity vector and stop circling instantaneously. The simple model is worse - a lot worse. It is a disservice to the soaring community to assert that the simple model is better because the more sophisticated model is somehow "impossible" - that is bunk. It might be "impossible" to program for someone who doesn't understand aircraft dynamics, but that's not the same as claiming that it is an impossible task. Aircraft have flown on autopilots that use the same basic principles for decades - including ones that can autoland under all kinds of non-deterministic airmass dynamics.

Closure rate is an example of a simple projection model - it is the simplest you can imagine as it understands NOTHING about the dynamics of the underlying systems. Simpler is not better in this case - it is quite useful to know that gliders have energy and stall speeds and minimum turn radii to understand whether a glider getting closer to you is actually on a path to hit you or not.

As an algorithm closure rate implicitly assumes the instantaneous closure rate will continue unabated until zero distance and collision. This is NOT true - hardly ever. In only a small percentage of cases to aircraft getting closer represent a collision threat. Without a more sophisticated algorithm you either have to warn the pilot of ALL of them (hardly useful) or set a threshold on closure rate and distance that delays the warning to reduce the false warnings. You would have to set it pretty tight in terms of time to impact to get the warnings down to a tolerable level and even so you will get a lot.

Thought experiment - imagine you are on a low, flat final glide at 70 kts and a bunch of other gliders are well above you and well to either side of you at 150 knots on fast final glides. None of them represent collision risks but they will have a very high rate of closure with a very short time to impact up until they get nearly abeam of you at which point the closure rate will fall of dramatically. Maybe it would be good not to panic you unless someone actually points their glider close enough to you to be a legitimate risk.

Note: As I understand the FLARM algorithm, it is basically indistinguishable from a closure rate algorithm at close range and for high closure rates because it becomes basically mathematically impossible for the paths not to intersect given the "dynamic radius" that FLARM puts on position estimates. So in the above scenario, if one of those gliders steepend their decent and started turing towards you you would get a warning.

So to summarize - it appears based on your statements that the T-Advisor does use a collision algorithm - it's just a much more simplistic one than FLARM does. This algorithm is based on closure rate and (if I understand you) distance. Depending on the thresholds set for issuing a warning this will generate either more false negatives or (more likely given your statement about 100% probability - a very loose standard since it assumes nothing about what the airplane is physically capable of doing) it will generate a lot more false positive warnings.

I'd be interested to see a cockpit video of one of these systems in flight in a situation with, say 30 other gliders milling about in a thermal in a pre-start situation at a contest.

9B

On Friday, May 29, 2015 at 9:43:32 PM UTC+2, Andy Blackburn wrote:
On Friday, May 29, 2015 at 2:13:45 AM UTC-7, Lucas wrote:
you are a control system engineer, therefore you work with deterministic systems. The glider pilot, whose actions depend on atmospheric conditions unknown a priori (i.e. vertical gusts/thermals), is not a deterministic system.

Nope - control systems engineers spend most of their time dealing with the effects of noise and non-determinist effects - no system is perfectly deterministic and you assume so at your peril. The issue is to figure out "how" deterministic a system is . Also, if you make NO assumptions about aircraft dynamics many systems are not accurately observable or controllable. Knowing what the aircraft can do dynamically makes a huge difference in knowing where it is and where it can possibly go. Assuming nothing about aircraft dynamics in your system IS making an assumption - you assume the system will keep doing exactly what it was doing when you measured it last - if it was going up it will keep going up indefinitely, if it was circling you assuming it will follow the instantaneous velocity vector and stop circling instantaneously. The simple model is worse - a lot worse. It is a disservice to the soaring community to assert that the simple model is better because the more sophisticated model is somehow "impossible" - that is bunk. It might be "impossible" to program for someone who doesn't understand aircraft dynamics, but that's not the same as claiming that it is an impossible task. Aircraft have flown on autopilots that use the same basic principles for decades - including ones that can autoland under all kinds of non-deterministic airmass dynamics.

Closure rate is an example of a simple projection model - it is the simplest you can imagine as it understands NOTHING about the dynamics of the underlying systems. Simpler is not better in this case - it is quite useful to know that gliders have energy and stall speeds and minimum turn radii to understand whether a glider getting closer to you is actually on a path to hit you or not.

As an algorithm closure rate implicitly assumes the instantaneous closure rate will continue unabated until zero distance and collision. This is NOT true - hardly ever. In only a small percentage of cases to aircraft getting closer represent a collision threat. Without a more sophisticated algorithm you either have to warn the pilot of ALL of them (hardly useful) or set a threshold on closure rate and distance that delays the warning to reduce the false warnings. You would have to set it pretty tight in terms of time to impact to get the warnings down to a tolerable level and even so you will get a lot.

Thought experiment - imagine you are on a low, flat final glide at 70 kts and a bunch of other gliders are well above you and well to either side of you at 150 knots on fast final glides. None of them represent collision risks but they will have a very high rate of closure with a very short time to impact up until they get nearly abeam of you at which point the closure rate will fall of dramatically. Maybe it would be good not to panic you unless someone actually points their glider close enough to you to be a legitimate risk.

Note: As I understand the FLARM algorithm, it is basically indistinguishable from a closure rate algorithm at close range and for high closure rates because it becomes basically mathematically impossible for the paths not to intersect given the "dynamic radius" that FLARM puts on position estimates. So in the above scenario, if one of those gliders steepend their decent and started turing towards you you would get a warning.

So to summarize - it appears based on your statements that the T-Advisor does use a collision algorithm - it's just a much more simplistic one than FLARM does. This algorithm is based on closure rate and (if I understand you) distance. Depending on the thresholds set for issuing a warning this will generate either more false negatives or (more likely given your statement about 100% probability - a very loose standard since it assumes nothing about what the airplane is physically capable of doing) it will generate a lot more false positive warnings.

I'd be interested to see a cockpit video of one of these systems in flight in a situation with, say 30 other gliders milling about in a thermal in a pre-start situation at a contest.

9B

Your note doesn't solve the basic problem: you cannot "predict" where the pilot will go in 10-20-... seconds.
Therefore any "prediction" is not possible.
This is very simple: there is no way of predicting what the pilot will do in most situations.
As already wrote, it's enough to look at a flight trace.
And, pardon me, but I have been working on flight dynamics for 19 years now.. I don't know about you, but the background seems quite different at this point.