Page 1 of 2 12 LastLast
Results 1 to 20 of 33

Thread: Dodge Neural Network inside

  1. #1
    Tuner in Training
    Join Date
    Nov 2018
    Posts
    15

    Dodge Neural Network inside

    Ever look at all those tricky numbers in the Neural Network settings and wonder "hmmmmm"...

    What's inside?

    VE table.png

    More on dealing with it to come


    -Deep Tuning

  2. #2
    Senior Tuner
    Join Date
    May 2015
    Posts
    1,864
    Hope you have some further info than we have already played around with. Also positive not everything is mapped out in any OS yet by HP, so aside from theory I don't think we have the ability to tune NN (not sure we should want to anyways).

  3. #3
    Tuner in Training
    Join Date
    Nov 2018
    Posts
    15
    Quote Originally Posted by 06300CSRT8 View Post
    Hope you have some further info than we have already played around with. Also positive not everything is mapped out in any OS yet by HP, so aside from theory I don't think we have the ability to tune NN (not sure we should want to anyways).
    I'd say there are a few OS's that aren't correct (at least with what I get for output from the NN from stuff in the Jeep Wrangler series), and I certainly don't think that an average person is going to sit down to fully tune a 5-Dimension fueling model from scratch.

    The good news is, I'm working on a way for the Dodge community to break away from disabling the NN and lying about target FA and injector data to get what they want.

    The other good news is: it's pretty close

    Hopefully, this will get some attention from the team at HPT so that we can further develop a way to implement a smooth strategy!

    -Deep Tuning

  4. #4
    Senior Tuner
    Join Date
    May 2015
    Posts
    1,864
    Good stuff...

    But just to correct your statement...when you turn off NN your injector data, target FA, and VE tables need to be spot on, accurate, and true. It is when you have NN ENABLED, you have to lie to the ECU with bad injector data to achieve the target FA you want. I would be nice to be able to input known good injector data and tweak the NN input layers to achieve the desired FA.

  5. #5
    Tuner in Training
    Join Date
    Nov 2018
    Posts
    15
    Yeah, oopsie! That was an and/or situation where it should have been a hard "or", but what's more important is we recognize this:

    There are inherit disadvantages in either school of thought for tuning these things.
    Either way, it's always bothered me like a splinter under the skin, "I'll make these adjustments, but I know it's not exactly right..."

    So I'm just hoping to offer a solution for the guys that want to do be able to adjust something more than just settling for "how everyone else does it", but I'm certainly not trying to make anyone change what works for them. In fact, I've had this network calculator for a little while now and I still make quick adjustments the old fashioned way because it's just plain easier in _a lot_ of cases. I'll add details to how it'll work in the upcoming days (without all the boring calculus and things).

    -Deep Tuning

  6. #6
    Tuner in Training
    Join Date
    May 2018
    Posts
    36
    I am interested

  7. #7
    Advanced Tuner
    Join Date
    Feb 2016
    Posts
    294
    If you rationalise the cam tables to suit performance rather than emissions, the VE becomes stable and NN is no longer required nor really desirable.
    You can get better power, drivability and economy than from the factory without using NN or MDS.
    But it sounds like a fun exercise.

  8. #8
    Advanced Tuner
    Join Date
    May 2016
    Posts
    966
    Quote Originally Posted by Hemituna View Post
    If you rationalise the cam tables to suit performance rather than emissions, the VE becomes stable and NN is no longer required nor really desirable.
    You can get better power, drivability and economy than from the factory without using NN or MDS.
    But it sounds like a fun exercise.
    I agree. Fix the cam timing so it's not moving the cams all over the place and fueling stays very stable with NN off and tune setup right.

  9. #9
    Tuner in Training
    Join Date
    Nov 2018
    Posts
    15
    But what if you could just let the cam be free OR freely change it and have an all-points-mapped configuration (as was the intention of all that crafty work by the OEM in the first place)? I mean ... I get it. Good enough is good enough, but isn't better also better ?

    I'll keep working on this so we can all find out how large this can of worms actually is. (I'm not sure that's grammatically correct, but it sounds okay in my head) I'll be doing some generalizations on training the system to predict what I want to see and as soon as it can reliably learn some fake crap, then we will trial on some small mod Hellcats and a few SC V6 platforms. Exciting

    -Deep Tuning

  10. #10
    Tuner in Training
    Join Date
    Nov 2018
    Posts
    15
    Also, I do appreciate the feedback from all of you guys! There are a few of you I consider a huge staple in the Dodge tuning community and my hats off to you.



    -Deep Tuning

  11. #11
    Quote Originally Posted by 06300CSRT8 View Post
    Hope you have some further info than we have already played around with. Also positive not everything is mapped out in any OS yet by HP, so aside from theory I don't think we have the ability to tune NN (not sure we should want to anyways).
    Which OS are you working with? Everything should be available on most OSes unless they haven't been updated yet.

    Quote Originally Posted by Deep Tuning View Post
    I'd say there are a few OS's that aren't correct (at least with what I get for output from the NN from stuff in the Jeep Wrangler series), and I certainly don't think that an average person is going to sit down to fully tune a 5-Dimension fueling model from scratch.

    The good news is, I'm working on a way for the Dodge community to break away from disabling the NN and lying about target FA and injector data to get what they want.

    The other good news is: it's pretty close

    Hopefully, this will get some attention from the team at HPT so that we can further develop a way to implement a smooth strategy!

    -Deep Tuning
    I may have dabbled with this before...
    L2l8WPc.png

    I'm translating the input and output values to their actual value vs you leaving them in neuron form, but same idea.
    My desire to make a calculator/trainer for this is present, there's just a lot of other tasks on my plate at the same time.
    If its not broke, just give it time.

  12. #12
    Tuner in Training
    Join Date
    Nov 2018
    Posts
    15
    Well the calculator part is already done on this side, but the backprop section and how to implement viable training data is still in the works. It's a fortune of time invested! But I'd love to talk about it if you have time to spare

    As far as the model that looked "broken" I can't say for certain, I'm sure I can find the file somewhere but if memory serves, it was a newer JL.


    -Deep Tuning

    p.s. you did way more breakpoints than I cared to do hahaha! I gave it 17x17 and said good enough. All in Excel on first set of calculators
    Last edited by Deep Tuning; 11-28-2018 at 09:53 PM.

  13. #13
    Quote Originally Posted by Deep Tuning View Post
    Well the calculator part is already done on this side, but the backprop section and how to implement viable training data is still in the works. It's a fortune of time invested! But I'd love to talk about it if you have time to spare

    As far as the model that looked "broken" I can't say for certain, I'm sure I can find the file somewhere but if memory serves, it was a newer JL.


    -Deep Tuning

    p.s. you did way more breakpoints than I cared to do hahaha! I gave it 17x17 and said good enough. All in Excel on first set of calculators
    Yeah, that's kind of where the implementing this reliably for people dropped off. Building a reliable and quick trainer for the network based on user data is tough (and makes the assumption the data collected is valid and good, which it may very well not be).
    Its not a crazy big network, so training it isn't necessarily hard but getting good data to train it is. You need to collect 5 pieces of data reliably and use that to train the network without overtraining.

    As to the size, I'm using MATLAB and the way I have it setup, the amount of breakpoints is an arbitrary number. I chose 100 because its a nice round number that takes no time at all to spit out a graph.
    More power to you if you tried this in Excel to start!
    If its not broke, just give it time.

  14. #14
    Advanced Tuner
    Join Date
    Feb 2016
    Posts
    294
    Quote Originally Posted by Deep Tuning View Post
    But what if you could just let the cam be free OR freely change it and have an all-points-mapped configuration (as was the intention of all that crafty work by the OEM in the first place)? I mean ... I get it. Good enough is good enough, but isn't better also better ?
    I'll keep working on this so we can all find out how large this can of worms actually is. (I'm not sure that's grammatically correct, but it sounds okay in my head) I'll be doing some generalizations on training the system to predict what I want to see and as soon as it can reliably learn some fake crap, then we will trial on some small mod Hellcats and a few SC V6 platforms. Exciting
    -Deep Tuning
    The blown stuff especially dont like the cam moving too much.
    A little is good...but anymore is purely for emissions and is to the detriment of performance.
    But knock yourself out.
    Will be interested as to where you end up on the blown pentastars as they perform best a mile from where the factory settings are.

  15. #15
    Tuner in Training
    Join Date
    Nov 2018
    Posts
    15
    Yeah, the pentastar engines were shocking, almost feels sinful having to change so much to get them running right.

  16. #16
    Tuner
    Join Date
    May 2007
    Location
    Seattle, WA
    Posts
    60
    I would love to learn more about getting this setup and going and pointers to where to start?
    2018 Jeep Trackhawk - Blue

    What we do in life, Echos in eternity

  17. #17
    Not to thread jack, but Since hemituna mentioned it, ive always felt the NN being off the car felt stronger. but my issue was always the cam swing and the VE model being off and dialing it in. My issue was figuring out how to rationalize the cam tables. The 5.7 6.4 ect only use the exhaust values right? so do you just make them match the WOT values?

  18. #18
    Tuner in Training
    Join Date
    Nov 2018
    Posts
    15
    Rationalizing the cam tables would just really mean limiting overall movement from table to table.

    More to the point would be making all the tables the same (or atleast similar). If one table is an extreme case from another (Normal table to Scavenging table for instance) we would want to make it more predictable that the cam would only get to certain position regardless of selected table.

    As a side note, both intake and exhaust positions are used for input to the NN even though its still "exhaust only" on the V8 models, but while running, the intake camshaft position is linked to the exhaust but is always the inverse position (monitor your Intake Cam and Exhaust Cam Input Neurons in the VCM Scanner software. For example, one might be 0.234 while the other is -0.234)

    This would make training (or retraining) the network a little more friendly due to the limited possibility of the deviation in dimensions. Since 4 inputs and 1 predicted output = 5 dimensions, the V8 engines really only link to 4 dimensions for airflow mapping. The word "friendly" above does not mean easy, but less friendly yet will be the full 5 dimensions on dual VVT engines (intake and exhaust independently controlled).

    I think I'll make a couple pictures to share how 5D mapping would work without the NN function! Excel chart time!

  19. #19
    Tuner in Training
    Join Date
    Nov 2018
    Posts
    15
    3D Mapping (easy - done fast)
    3D VE.png

    4D Mapping (one cam moves - takes longer but can be done with patience)
    4D VE.png

    5D Mapping (game over... I'm going back to Carburetors!)
    5D VE.png

    -Deep Tuning

  20. #20
    Tuner in Training
    Join Date
    Nov 2018
    Posts
    15
    Here's something else I wanted to share. Should do this as a video, but don't want to give too much hard work away at the moment!

    Standard VE model with cam angles at some values:
    standard model at cam angles x-y.png

    Modification to ONE weight in the first layer:
    VE - L1W2delta.png

    Modification to ONE weight in the second layer:
    VE - L2W9delta.png

    So... initial concept has been applied in Excel for my own need for visualization and understanding, but the intent is to upscale and streamline a process that handles the conversion of weights, biases, and perhaps training data where we have added a turbo or SC and 2/3 bar MAP sensor or installed a camshaft phase limiter. Ideally, the user should be able to modify a graph or table of numbers/axis values, and the backend of the system will modify the required areas to best match user desire. Dream big, right!

    -Deep Tuning