JSBSim's randomseed impact on turbulence generation

[Dobid]

Hello JSBSim Team,

I use the JSBSim python library version 1.2.0 on a Ubuntu 22.04 machine.
I have simplified my use case as such. I want to simulate 2 episodes of 2000 timesteps each of flight, with placeholder fcs commands set to trim.

• Experiment 1

  • Episode 1:
    I 1) set constant wind 2) turbulence parameters 3) the random seed to 1.
    I reset to initial conditions.
  • Episode 2:
    I 1) set constant wind 2) turbulence parameters 3) the random seed to 2.
    Which gives me this figure
    

• Experiment 2
  Then I simulate the same experiment only swapping the seeds for both episodes: (episode 1, seed=1; episode 2, seed=2) --> (episode 1, seed=2; episode 2, seed=1) and plot the following figure:

As you can see, the wind speed figures for both experiments are not identical when looking at the same seeds. The profile of the curves looks the same, but there are offsets. I would expect identical seeds to produce the exact same turbulent wind across experiments.

Could you please provide an explanation ?

In case you need it, here is the python script I wrote, I hope it's minimal enough for you to have a look.

import numpy as np
import jsbsim
import matplotlib.pyplot as plt
from fw_jsbgym.trim.trim_point import TrimPoint


def set_wind(fdm):
    # Setting up the wind: 5kph constant wind and severe turbulence
    fdm["atmosphere/wind-north-fps"] = 5 * 0.3048
    fdm["atmosphere/wind-east-fps"] = 0
    fdm["atmosphere/wind-down-fps"] = 0
    fdm["atmosphere/turb-type"] = 3
    fdm["atmosphere/turbulence/milspec/windspeed_at_20ft_AGL-fps"] = 75
    fdm["atmosphere/turbulence/milspec/severity"] = 6


def main():
    # Setting up the JSBSim FDM
    fdm = jsbsim.FGFDMExec("fdm_descriptions")
    fdm.set_debug_level(0)
    fdm.load_model("x8")
    fdm.load_ic("initial_conditions/x8_basic_ic.xml", False)
    fdm.run_ic()

    # 2 episodes of 2000 timesteps each
    episode_max_tsteps = 2000
    n_episodes = 2
    max_tsteps = episode_max_tsteps * n_episodes

    # One seed per episode
    seeds = [1, 2] # [1, 2] for 1st plot, [2, 1] for 2nd plot
    fdm["simulation/randomseed"] = seeds[0]

    set_wind(fdm)

    # Setting up the commands to be at trim
    trim_cmd = TrimPoint("x8")

    # Create storing buffers
    roll = np.zeros((n_episodes, episode_max_tsteps))
    pitch = np.zeros((n_episodes, episode_max_tsteps))
    yaw = np.zeros((n_episodes, episode_max_tsteps))
    wind_n = np.zeros((n_episodes, episode_max_tsteps))
    wind_e = np.zeros((n_episodes, episode_max_tsteps))
    wind_d = np.zeros((n_episodes, episode_max_tsteps))

    total_t = 0
    ep_cnt = 0
    ep_t = 0
    while total_t < max_tsteps:
        # input commands
        fdm["fcs/aileron-cmd-norm"] = trim_cmd.aileron # -4.288e-15 / 0.0
        fdm["fcs/elevator-cmd-norm"] = trim_cmd.elevator # 0.0406
        fdm["fcs/throttle-cmd-norm"] = trim_cmd.throttle # 0.1295
        # print(f"fdm['simulation/randomseed'] = {fdm['simulation/randomseed']}")

        # get the attitude of the aircraft
        roll[ep_cnt, ep_t] = fdm["attitude/roll-rad"]
        pitch[ep_cnt, ep_t] = fdm["attitude/pitch-rad"]
        yaw[ep_cnt, ep_t] = fdm["attitude/psi-rad"]

        # get the wind speeds
        wind_n[ep_cnt, ep_t] = fdm["atmosphere/total-wind-north-fps"]
        wind_e[ep_cnt, ep_t] = fdm["atmosphere/total-wind-east-fps"]
        wind_d[ep_cnt, ep_t] = fdm["atmosphere/total-wind-down-fps"]
        fdm.run()

        total_t += 1
        ep_t += 1
        if (ep_t == episode_max_tsteps) and (ep_cnt < n_episodes - 1):
            ep_cnt += 1
            ep_t = 0
            fdm.reset_to_initial_conditions(0)
            fdm["simulation/randomseed"] = seeds[ep_cnt]
            set_wind(fdm)
            print(f"Resetting to IC with seed {seeds[ep_cnt]}")

    # Plot attitude and windspeed, episode 0 and 1 are side by side
    fig, axs = plt.subplots(2, 2)
    # Episode 0
    axs[0, 0].plot(roll[0, :], label="roll")
    axs[0, 0].plot(pitch[0, :], label="pitch")
    axs[0, 0].plot(yaw[0, :], label="yaw")
    axs[0, 0].legend()
    axs[0, 0].grid()
    axs[0, 0].set_title(f"Attitude Angles [rad] - EP1 seed {seeds[0]}")

    axs[1, 0].plot(wind_n[0, :], label="wind_n")
    axs[1, 0].plot(wind_e[0, :], label="wind_e")
    axs[1, 0].plot(wind_d[0, :], label="wind_d")
    axs[1, 0].legend()
    axs[1, 0].grid()
    axs[1, 0].set_title(f"Wind Speeds[rad] - EP1 seed {seeds[0]}")

    # Episode 1
    axs[0, 1].plot(roll[1, :], label="roll")
    axs[0, 1].plot(pitch[1, :], label="pitch")
    axs[0, 1].plot(yaw[1, :], label="yaw")
    axs[0, 1].legend()
    axs[0, 1].grid()
    axs[0, 1].set_title(f"Attitude Angles[rad] - EP2 seed {seeds[1]}")

    axs[1, 1].plot(wind_n[1, :], label="wind_n")
    axs[1, 1].plot(wind_e[1, :], label="wind_e")
    axs[1, 1].plot(wind_d[1, :], label="wind_d")
    axs[1, 1].legend()
    axs[1, 1].grid()
    axs[1, 1].set_title(f"Wind Speeds[rad] - EP2 seed {seeds[1]}")

    # plt.savefig('diff_seeds.png')
    plt.show()


if __name__ == '__main__':
    main()

Thank you in advance!

[seanmcleod70]

Quick initial look.

jsbsim/src/FGFDMExec.cpp

Line 170 in e1ea4e8

instance->Tie("simulation/randomseed", this, (iPMF)&FGFDMExec::SRand, &FGFDMExec::SRand);

jsbsim/src/FGFDMExec.cpp

Lines 1372 to 1376 in e1ea4e8

	void FGFDMExec::SRand(int sr)
	{
	RandomSeed = sr;
	RandomGenerator->seed(RandomSeed);
	}

jsbsim/src/FGFDMExec.cpp

Lines 80 to 81 in e1ea4e8

	FGFDMExec::FGFDMExec(FGPropertyManager* root, std::shared_ptr<unsigned int> fdmctr)
	: RandomSeed(0), RandomGenerator(make_shared<RandomNumberGenerator>(RandomSeed)),

jsbsim/src/FGFDMExec.h

Lines 692 to 693 in e1ea4e8

	unsigned int RandomSeed;
	std::shared_ptr<RandomNumberGenerator> RandomGenerator;

jsbsim/src/FGFDMExec.h

Line 630 in e1ea4e8

auto GetRandomGenerator(void) const { return RandomGenerator; }

jsbsim/src/models/atmosphere/FGWinds.cpp

Lines 75 to 76 in e1ea4e8

	FGWinds::FGWinds(FGFDMExec* fdmex)
	: FGModel(fdmex), generator(fdmex->GetRandomGenerator())

jsbsim/src/models/atmosphere/FGWinds.h

Line 397 in e1ea4e8

std::shared_ptr<RandomNumberGenerator> generator;

jsbsim/src/models/atmosphere/FGWinds.cpp

Lines 304 to 307 in e1ea4e8

	nu_u = generator->GetNormalRandomNumber(),
	nu_v = generator->GetNormalRandomNumber(),
	nu_w = generator->GetNormalRandomNumber(),
	nu_p = generator->GetNormalRandomNumber(),

So it appears as if FGWinds gets a pointer to RandomGenerator from FGFDMExec which is updated with the seed when you set simulation/randomseed, but maybe another instance is being created rather than getting a pointer to the original one? Will need to double-check.

But even if a pointer to the original FGFDMExec is being used by FGWinds there are other modules/systems that also make use of it. So isn't there some chance that depending on other conditions in the sim there may be the case that in some runs the order of calls to the random number generator isn't consistent?

[bcoconni]

But even if a pointer to the original FGFDMExec is being used by FGWinds there are other modules/systems that also make use of it. So isn't there some chance that depending on other conditions in the sim there may be the case that in some runs the order of calls to the random number generator isn't consistent?

Yep ! And @Dobid, I see that you call the function TrimPoint at the start:

    fdm["simulation/randomseed"] = seeds[0]

    set_wind(fdm)

    # Setting up the commands to be at trim
    trim_cmd = TrimPoint("x8")

but you don't call it after resetting to initial conditions:

            fdm.reset_to_initial_conditions(0)
            fdm["simulation/randomseed"] = seeds[ep_cnt]
            set_wind(fdm)
            print(f"Resetting to IC with seed {seeds[ep_cnt]}")

As @seanmcleod70 mentioned you need to have the exact same sequence of events between the 2 episodes to obtain identical behavior between your tests. And by "exact same sequence", I mean strictly the same, any deviation whatsoever will lead to different results. That's the point of randomness 😉

[Dobid]

TrimPoint isn't a function it's just a class I coded that stores the values of aileron, elevator and throttle commands to achieve a trimmed flight for my x8 UAV. So they need to be set once at the beginning at the script since I don't need to change those commands during an episode, or from episode 1 to episode 2.

[seanmcleod70]

So I tried a very quick test.

Used the 737 cruise script and set the randomseed.

	<use aircraft="737" initialize="cruise_init"/>

	<run start="0" end="100" dt="0.008333">

		<property value="0"> simulation/notify-time-trigger </property>
		<property value="1"> simulation/randomseed </property>

Then set some breakpoints in JSBSim to take a look at the random number generator, and confirmed that FGWinds is getting a shared pointer to the one from FGFDMExec.

Then added this test call in FGWinds::Run to get a random number on each call.

bool FGWinds::Run(bool Holding)
{
  if (FGModel::Run(Holding)) return true;
  if (Holding) return false;

  // Testing
  double ran = generator->GetNormalRandomNumber();

And confirmed that I get consistent random numbers based on the seed, running each seed case a couple of times.

Seed 1:
-1.4028727091279207
-0.54974617895544953
-1.0451468104420230

Seed 2:
0.66244780319854357
-0.022564746291165897
-0.46089239739463156

So at least in this very limited test case the random seed functionality of JSBSim appears to be working as expected.

[seanmcleod70]

@Dobid so I was able to get consistent winds based on a randomseed of 1 and 2 using this code. Running it multiple times while changing the seed and comparing the outputs.

import jsbsim
import matplotlib.pyplot as plt

fdm = jsbsim.FGFDMExec('..\\') 

fdm["simulation/randomseed"] = 1

fdm.load_model('737') # Load the aircraft 737

# Set engines running
fdm['propulsion/engine[0]/set-running'] = 1
fdm['propulsion/engine[1]/set-running'] = 1

fdm['ic/h-sl-ft'] = 20000
fdm['ic/vc-kts'] = 250
fdm['ic/gamma-deg'] = 0

fdm.run_ic()

fdm['simulation/do_simple_trim'] = 1

fdm["atmosphere/wind-north-fps"] = 5 * 0.3048
fdm["atmosphere/wind-east-fps"] = 0
fdm["atmosphere/wind-down-fps"] = 0
fdm["atmosphere/turb-type"] = 3
fdm["atmosphere/turbulence/milspec/windspeed_at_20ft_AGL-fps"] = 75
fdm["atmosphere/turbulence/milspec/severity"] = 6

wind_n = []
wind_e = []
wind_d = []

while fdm.get_sim_time() < 10:
    fdm.run()

    wind_n.append(fdm["atmosphere/total-wind-north-fps"])
    wind_e.append(fdm["atmosphere/total-wind-east-fps"])
    wind_d.append(fdm["atmosphere/total-wind-down-fps"])

plt.figure()

plt.plot(wind_n)
plt.plot(wind_e)
plt.plot(wind_d)

plt.show()

Seed = 1

Seed = 2

[Dobid]

Well, I've investigated a bit and I don't think that the random seed is the problem. Because in my original post, the same seed (on the two different experiments) gives the same wind profile (the same "jittering") the only thing making both winds different (for example seed 2 for EXP1&2) seems to be some offset.

Now if I take your script with the 737 and if I ask to disable turbulence and set wind speed back to 0 after 5 seconds, I obtain this:

So, now my question is: Why is the constant wind not 0 after I set the wind properties to 0 ?

This seems to be the cause of the offset I'm seeing in my experiments with the x8.

Here's what I changed in the code you provided:

while fdm.get_sim_time() < 10:
    fdm.run()

    wind_n.append(fdm["atmosphere/total-wind-north-fps"])
    wind_e.append(fdm["atmosphere/total-wind-east-fps"])
    wind_d.append(fdm["atmosphere/total-wind-down-fps"])

    # Resetting wind to 0 kph and no turbulence after 5 seconds
    if fdm.get_sim_time() > 5.0:
        fdm["atmosphere/wind-north-fps"] = 0
        fdm["atmosphere/wind-east-fps"] = 0
        fdm["atmosphere/wind-down-fps"] = 0
        fdm["atmosphere/turb-type"] = 0
        fdm["atmosphere/turbulence/milspec/windspeed_at_20ft_AGL-fps"] = 0
        fdm["atmosphere/turbulence/milspec/severity"] = 0

[seanmcleod70]

@bcoconni I'm wondering whether we shouldn't provide an option to specify an independent seed for FGWinds, independent of the seed used by FGFDMExec.

Let's say I'm wanting to test an autoland autopilot and how well it performs with turbulence, and so I need to compare how different versions of it perform against the exact same set of turbulence.

Now let's assume different versions of the autoland autopilot make use of a different number of FGSensors, or maybe the same set, but in some cases I want to compare how the autoland performs when the sensor has no noise at all versus with noise modelled.

Given that FGSensor makes use of the same instance of the random number generator it means now that there would be cases where I wouldn't be able to compare the performance of different autloland implementations against the exact same set of winds/turbulence.

jsbsim/src/models/flight_control/FGSensor.cpp

Lines 54 to 55 in e1ea4e8

	FGSensor::FGSensor(FGFCS* fcs, Element* element)
	: FGFCSComponent(fcs, element), generator(fcs->GetExec()->GetRandomGenerator())

jsbsim/src/models/flight_control/FGSensor.cpp

Line 167 in e1ea4e8

if (noise_variance != 0.0) Noise(); // models noise

jsbsim/src/models/flight_control/FGSensor.cpp

Lines 185 to 192 in e1ea4e8

	void FGSensor::Noise(void)
	{
	double random_value=0.0;
	if (DistributionType == eUniform)
	random_value = generator->GetUniformRandomNumber();
	else
	random_value = generator->GetNormalRandomNumber();

[bcoconni]

Let's say I'm wanting to test an autoland autopilot and how well it performs with turbulence, and so I need to compare how different versions of it perform against the exact same set of turbulence.

That seems reasonable provided that we keep backward compatibility. In that purpose, the property simulation/randomseed could propagate its value/seed to all the random number generators (RNG) and the RNG of each class FGWinds and FGSensor can have their own property to set their seed.

There is also the question of whether each instance of FGSensor should have its own RNG ? Or if they will all share one common RNG ? The answer to this question will drive the implementation knowing that static variables are prohibited per #666 😉

Given that FGSensor makes use of the same instance of the random number generator

Yes, commit e27d0ae has kept the original intent which was to have a RNG that is common to all classes.

[Dobid]

I agree to this. I'm specifically in the case where I need to compare different autopilot controllers against the same wind conditions.

[seanmcleod70]

That seems reasonable provided that we keep backward compatibility.

Yep, the idea would be remain backwards compatible. So if no wind specific seed is specified the winds will continue to use the random number generator and it's seed from FGFDMExec. But if a wind specific seed is specified via some new property then the wind class would generate it's own instance of a random number generator with the supplied seed etc.

[seanmcleod70]

@Dobid it looks there is some side-effect of enabling and then disabling the turbulence.

If I remove the turbulence code then the wind values behave as expected. So will need to dig into the turbulence code to figure out what is going wrong when you try and disable the turbulence.

fdm["atmosphere/wind-north-fps"] = 5 * 0.3048
fdm["atmosphere/wind-east-fps"] = 3 * 0.3048
fdm["atmosphere/wind-down-fps"] = 2 * 0.3048
#fdm["atmosphere/turb-type"] = 3
#fdm["atmosphere/turbulence/milspec/windspeed_at_20ft_AGL-fps"] = 75
#fdm["atmosphere/turbulence/milspec/severity"] = 6
....
    if fdm.get_sim_time() > 5.0:
        fdm["atmosphere/wind-north-fps"] = 1
        fdm["atmosphere/wind-east-fps"] = 2
        fdm["atmosphere/wind-down-fps"] = 3

[seanmcleod70]

At first glance, I'll double-check later, my guess is that when the turbulence is switched off, the Turbulence() method on line 146 is no longer called, but the vector vTurbulenceNED isn't zeroed out, rather it's left at it's last value it was assigned by the last call to Turbulence(). In which case it'll be a simple fix.

jsbsim/src/models/atmosphere/FGWinds.cpp

Lines 141 to 149 in e1ea4e8

	bool FGWinds::Run(bool Holding)
	{
	if (FGModel::Run(Holding)) return true;
	if (Holding) return false;
	if (turbType != ttNone) Turbulence(in.AltitudeASL);
	if (oneMinusCosineGust.gustProfile.Running) CosineGust();
	vTotalWindNED = vWindNED + vGustNED + vCosineGust + vTurbulenceNED;

Will need to double-check if similar issues may affect vGustNED and vCosineGust.

[seanmcleod70]

I've checked the source code for vGustNED and VCosineGust and tested them out and they don't have the same issue as vTurbulenceNED. I'll submit a PR to fix the issue with vTurbulenceNED.

[Dobid]

Great to hear ! Sorry for the dumb question, but how can I update the jsbsim python package on my computer once the PR has been accepted ? I am using conda btw.

[seanmcleod70]

Hmm, I think if you want to immediately update your jsbsim python package you would need to build it locally after the PR is submitted.

Otherwise wait for the next jsbsim release - https://github.com/JSBSim-Team/jsbsim/releases

When a new release comes out the python package at PyPi and Conda Forge are updated:

https://pypi.org/project/jsbsim/#history

https://anaconda.org/conda-forge/jsbsim

[Dobid]

Hmm, the next release isn't scheduled to be in the upcoming weeks or month I suppose ? So I will build the python package then !
Also, I second your suggestion about having a separate random number generator for the wind, I would make my life a lot easier, so let me know if you want/need me to open an issue, so you guys can track it more easily !

[seanmcleod70]

@Dobid yep it could be a while before the next release.

I've just submitted a pull request - #1252 to fix the issue where the turbulence effect isn't removed completely when turbulence is disabled.

I'll submit a separate pull request for separate random number generator work.

[Dobid]

Cool, thank you !

[seanmcleod70]

I've written some code to allow a separate random seed to be specified for FGWinds.

However, I'm battling to write a test case for it using our python test framework. Here is my attempt so far. For now I'm testing with simulation/randomseed being the same in both runs since I haven't rebuilt the jsbsim python module yet for the new FGWinds seed.

from JSBSim_utils import JSBSimTestCase, CreateFDM, RunTest


class TestTurbulenceRandomSeed(JSBSimTestCase):

    def testTurbulenceRandomSeed(self):
        # Test that the wind turbulence random seed is reproducible
        wind_random_seed = 2
        wn1, we1, wd1 = self.captureTurbulence(wind_random_seed, 5)
        wn2, we2, wd2 = self.captureTurbulence(wind_random_seed, 5)
        for i in range(len(wn1)):
            self.assertAlmostEqual(wn1[i], wn2[i], delta=1E-8)
            self.assertAlmostEqual(we1[i], we2[i], delta=1E-8)
            self.assertAlmostEqual(wd1[i], wd2[i], delta=1E-8)

    def captureTurbulence(self, wind_seed, exec_seed):
        fdm = self.create_fdm()

        # Set random seeds for FGFDMExec and FGWinds
        fdm["simulation/randomseed"] = exec_seed
        fdm["atmosphere/randomseed"] = wind_seed

        fdm.load_model('A4') 

        # Set engine running
        fdm['propulsion/engine[0]/set-running'] = 1

        fdm['ic/h-sl-ft'] = 20000
        fdm['ic/vc-kts'] = 250
        fdm['ic/gamma-deg'] = 0

        fdm.run_ic()

        fdm['simulation/do_simple_trim'] = 1

        # Setup turbulence
        fdm["atmosphere/turb-type"] = 3
        fdm["atmosphere/turbulence/milspec/windspeed_at_20ft_AGL-fps"] = 75
        fdm["atmosphere/turbulence/milspec/severity"] = 6

        wn = []
        we = []
        wd = []

        while fdm.get_sim_time() < 0.1:
            fdm.run()
            wn.append(fdm['atmosphere/total-wind-north-fps'])
            we.append(fdm['atmosphere/total-wind-east-fps'])
            wd.append(fdm['atmosphere/total-wind-down-fps'])

        # Attempted cleanup to force new FDM for next execution, not helping...
        self.delete_fdm()
        del fdm
        fdm = None

        print('')
        print(wn)

        return (wn, we, wd)

RunTest(TestTurbulenceRandomSeed)

However, what I'm seeing is that creating 2 separate instances of the FDM via calls to captureTurbulence() doesn't result in the same sequence of random numbers.

(base) C:\source\jsbsim\tests>python .\TestTurbulenceRandomSeed.py
testTurbulenceRandomSeed (__main__.TestTurbulenceRandomSeed) ...
[0.047354487544379376, 3.460354080430806, 4.636633740213811, -0.4374137958205208, -0.06377215234821773, -1.5456835409458707, -3.4411298921520492, -6.057946773083572, -5.709936753084856, -5.628235304626574, -6.480877261842454, -5.210906810642065, -2.2376367571469964]

[-2.1842398652974198, 1.2417689243655539, 2.426390607718044, -2.652099515681857, -2.2717034356215526, -3.750613502555642, -5.643749459388144, -8.259702814605022, -7.904921571924209, -7.81701415925393, -8.665327778342736, -7.3868098829003275, -4.401647923547139]
FAIL

======================================================================
FAIL: testTurbulenceRandomSeed (__main__.TestTurbulenceRandomSeed)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "C:\source\jsbsim\tests\TestTurbulenceRandomSeed.py", line 32, in testTurbulenceRandomSeed
    self.assertAlmostEqual(wn1[i], wn2[i], delta=1E-8)
AssertionError: 0.047354487544379376 != -2.1842398652974198 within 1e-08 delta (2.231594352841799 difference)

----------------------------------------------------------------------
Ran 1 test in 0.021s

FAILED (failures=1)
(base) C:\source\jsbsim\tests>

ANOTHER RUN:

(base) C:\source\jsbsim\tests>python .\TestTurbulenceRandomSeed.py
testTurbulenceRandomSeed (__main__.TestTurbulenceRandomSeed) ...
[0.047354487544379376, 3.460354080430806, 4.636633740213811, -0.4374137958205208, -0.06377215234821773, -1.5456835409458707, -3.4411298921520492, -6.057946773083572, -5.709936753084856, -5.628235304626574, -6.480877261842454, -5.210906810642065, -2.2376367571469964]

[-2.1842398652974198, 1.2417689243655539, 2.426390607718044, -2.652099515681857, -2.2717034356215526, -3.750613502555642, -5.643749459388144, -8.259702814605022, -7.904921571924209, -7.81701415925393, -8.665327778342736, -7.3868098829003275, -4.401647923547139]
FAIL

======================================================================
FAIL: testTurbulenceRandomSeed (__main__.TestTurbulenceRandomSeed)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "C:\source\jsbsim\tests\TestTurbulenceRandomSeed.py", line 32, in testTurbulenceRandomSeed
    self.assertAlmostEqual(wn1[i], wn2[i], delta=1E-8)
AssertionError: 0.047354487544379376 != -2.1842398652974198 within 1e-08 delta (2.231594352841799 difference)

----------------------------------------------------------------------
Ran 1 test in 0.023s

FAILED (failures=1)
(base) C:\source\jsbsim\tests>

My guess is that given we're creating 2 instances of the FDM within the same process there is some state in the random number generation process that isn't being reset completely.

    /// Specify a new seed and reinitialize the random generation process.
    void seed(unsigned int value) {
      generator.seed(value);
      uniform_random.reset();
      normal_random.reset();
    }

[seanmcleod70]

Hmm, ignore my comment/guess about it being a single process issue. Here is a simple single process example where the random number generation does work as expected.

#include <iostream>
#include <random>
 
int main()
{
    //std::mt19937 gen;
    std::default_random_engine gen;
 
    std::normal_distribution<double> normal_random;
    
    for(int i=0; i<3; i++)
    {
        // Seed the engine with an unsigned int
        gen.seed(1);
        std::cout << "after seed by 1: " << gen() << '\n';
 
        normal_random.reset();
 
        for(int j=0; j<3; j++)
        {
            std::cout << normal_random(gen) << '\n';
        }
    }
}

after seed by 1: 16807
0.207042
1.61506
0.118772
after seed by 1: 16807
0.207042
1.61506
0.118772
after seed by 1: 16807
0.207042
1.61506
0.118772

[seanmcleod70]

@Dobid I've submitted a pull request - #1253 to allow a user to specify a separate random seed for wind turbulence.

[seanmcleod70]

I've been able to reproduce the issue with 2 instances of FGFDMExec being run serially in a single process.

void run(char* script)
{
  JSBSim::FGFDMExec* FDMExec = new JSBSim::FGFDMExec();
  FDMExec->LoadScript(SGPath(script));
  FDMExec->RunIC();
  bool result = true;
  while (result) result = FDMExec->Run();

  delete FDMExec;
}

int main(int argc, char* argv[])
{
  run("scripts\\random1.xml");
  run("scripts\\random2.xml");

  return 0;
}

With scripts random1.xml.

  <run start="0" end="15" dt="0.008333">

	  <property value="1"> simulation/randomseed </property>
	  <property value="3"> atmosphere/randomseed </property>

	  <property value="3"> atmosphere/turb-type </property>
	  <property value="75"> atmosphere/turbulence/milspec/windspeed_at_20ft_AGL-fps </property>
	  <property value="3"> atmosphere/turbulence/milspec/severity </property>	  

And script random2.xml.

  <run start="0" end="15" dt="0.008333">

	  <property value="2"> simulation/randomseed </property>
	  <property value="3"> atmosphere/randomseed </property>

	  <property value="3"> atmosphere/turb-type </property>
	  <property value="75"> atmosphere/turbulence/milspec/windspeed_at_20ft_AGL-fps </property>
	  <property value="3"> atmosphere/turbulence/milspec/severity </property>	  

[seanmcleod70]

@bcoconni it looks like we've been bitten by static issues again. So the random number generator is behaving as expected, but the second instance of FGFDMExec etc. when initialising in the same process isn't setting these to 0, so we get different wind results even with the same random number generator seed.

jsbsim/src/models/atmosphere/FGWinds.cpp

Lines 286 to 293 in 614a5b7

	// keep values from last timesteps
	// TODO maybe use deque?
	static double
	xi_u_km1 = 0, nu_u_km1 = 0,
	xi_v_km1 = 0, xi_v_km2 = 0, nu_v_km1 = 0, nu_v_km2 = 0,
	xi_w_km1 = 0, xi_w_km2 = 0, nu_w_km1 = 0, nu_w_km2 = 0,
	xi_p_km1 = 0, nu_p_km1 = 0,
	xi_q_km1 = 0, xi_r_km1 = 0;

[seanmcleod70]

I've pushed some more commits to PR - #1253 to fix the static local variables issue and also added the python test program that I couldn't originally get to work due to this static local variable issue.

[bcoconni]

The PR is now merged.

[seanmcleod70]

@Dobid see pull request - #1254 which now also allows a user to specify individual randomseeds for individual sensors. So you have control over the randomseed for turbulence and sensors now.