Well I have absolutely no problem saying this is way over my head...haha
Well I have absolutely no problem saying this is way over my head...haha
Deep Tuning,
If you need a programmer to assist, I'd love to learn a little bit more about how this all works. I'm not super familiar with neural network training, but I'm a software engineer by trade and love to learn
2018 Jeep Trackhawk - Blue
What we do in life, Echos in eternity
The joys of setting hyperparameters to train a neural network:
image.png
This was one of the earlier sessions that ended up "exploding" (no engines were harmed as this is all still in experimental phase!)
-Deep Tuning
Still ironing out details, but I thought I'd share another picture instead of 1000 words:
Training table error on prediction - single case.PNG
The training system works on broad generalizations at the moment while losing a little in the fine detail, but it's pretty damn close
-Deep Tuning
I am with Jay...not sure what you are showing here and too lazy to search...but are you suggesting that we can tune by changing NN parameters? When it comes to programming my skills are closer to being able to change the time a my microwave and furthest from hacking the pentagon.
Last edited by Homer; 12-21-2018 at 02:09 PM.
Microwave clock setting skills are lacking on this side. I'd rather not know what the time or date is!
And, I was just showing a difference map on a standard Hellcat predicted NN VE table vs. what I've been able to generate with training a completely randomized Neural Network. It's always fun trying to find out which methods were used for training, how much data is needed to train the system on, and then how to effectively change the outcome to suit a tuning requirement.
At the moment, I'm sending nearly 20,000 individual data points through the program. Each iteration takes about half a second to push through, get a prediction, measure the error, push the error backwards through the network to systematically update the weights and biases (which were randomly chosen initially). Then, the system gets to see the data again, predict an output, measure the error, and "fix" the error. Again.... and again... and again...
So yeah... The NN is tunable, but not quickly and not without some effort. Hopefully our work here will lead to a solution that is both easy and intuitive for the end user, and doesn't require a programming background to use.
-Deep Tuning
I spent some time over the holiday break to figure all this out and now I can now comfortably manipulate the NN to achieve new desirable results than what the stock NN outputs. Unfortunately, the the process I developed for myself isn't user friendly yet either. I went ahead and created this animation of what the ANN outputs from a Hellcat tune to help us all visualize what's going on in there.
Thanks DT for the inspiration!
Where did this end up?
Sorry guys,
Been dealing with life so took a small break after the holidays. Just as an update, I'm leaning towards two things.
1) This is NOT a quick solution to tuning (nor do I think it will be even on the fasted GPU machines). The fastest I've been able to get high accuracy numbers from scratch is about 15 minutes of training time. I don't think anyone trying to tune a car would want to wait this long in a customers car each time they wanted to make a small NN VE tweak.
2) after playing with disabling the NN on many cars for comparing notes, the tuning process is pretty damn fast. This is still the fastest/easiest approach.
I may not continue with the end user program, but I do have an Excel file I've built that allows me to see the effect of each weight and bias on the system output. It's been easy enough to use for small tweaks in cars with exceptions on systems with missing tables and several Jeep Wranglers where the weights and bias yield stupid looking tables.
-Deep Tuning
The tuning school teaches ypu in its course how to tune the ANN to make it perform the way you want. I have just gone through there process. While it doesnt go into the layers and such I believe that will probably be in there level 2 training. But i have ANN on getting anywhere from 12.2 - 11.9 AFR at WOT and staying near perfect 14.7 during normal driving like it should. Trims really dont engage much maybe to 7 percent as I got tired of fine tuning it after a full day. Im going to tune the VE Tables when I have some time but that will be after my move to virginia at end of march.
Well I'm glad someone is taking time to learn and teach the Dodge platform! Kudos to them for putting something together, and to you for investing time and money to learn.
I'm curious as to the approach, so I'll have to look up what they offer. Good luck with the move!
-Deep Tuning
Anyone ever tried copying newer 6.4 NN weights & biases to run 6.4 on an older ngc4 pcm? Would it work? Asking for a friend...
The answer is kinda obvious, but I'd still ask - did you end up with some tool or a method that could be used to adjust the NN in non-stock applications?
Have a question....I have had my motor tuned but it wasn't a great tune.....bought the HP tuner and started looking around and see ANN disabled......well tune is sorta close...coould just turning ANN back on let her make adjustments from there? Would this work??
Any updates?
When you say rationalize are you talking about removing the part throttle portion of the camshaft tables that create natural EGR & reduce pumping loses? I would imagine that the whole RPM column would have to be the same VVT value to achieve a stable VE. Therefore you can't have varying angles for different airflows in the same RPM column?