Feature Idea: Cell Color Using Confidence Interval

[CCS86]

After using Scanner for a short time now, I think that this feature would provide huge value.

One of the most critical parts to successful tuning is deciding what data is good and what is junk. Features like filtering help to weed out errant data, but it doesn't help to determine how scattered or precise the remaining data is, for a given histogram chart.

Example: You are using STFTs to tune some table. You map out your axes and labels to match some ECM table and populate with STFTs. It might be tempting to just copy those averages and apply them to correct the table. But, if the trim data is all over the place, there isn't a strong correlation to justify editing the table.

If you toggle between maximum and minimum values, and see a max of +25% and a min of -29%, it's probably best to be suspicious that the average of +5% might not be very significant. In reality, 90% of your cell data might be within a few percent of that +5% average, with only a few outliers giving you the large max/min. Or your data could be mostly over +20% and under -20%. You really need to use some type of statistical analysis to determine the confidence interval of your data:

https://en.wikipedia.org/wiki/Confidence_interval

If Scanner allowed you to set a range, let's say +/- 4%, it could calculate the percentage of the data which falls within +/-4% of the average, and then color the cell based on that calculated value.

Maybe you say that if 80% of that cell data falls within +/-4% of average, we color it green; and if only 20% of the data is within +/-4%, we color it red.

Being able to set up histogram charts like this would allow you to plot fuel trims (or other relevant PIDs) against a number of different parameters (MAF period, fuel flow rate, fuel load, temps, etc) and see which has better correlation, guiding you towards what needs to be tuned the most.

Thoughts?

[CCS86]

I honestly though there would be more interest in this idea.

Without something like this, you are really just guessing at where to apply corrections, and using trial and error.

Maybe a visual will help:

In these two cases, the average, min, and max values reported by scanner will be exactly the same. But it is pretty clear to see the top data set has a high level of confidence (good correlation), and the bottom data set is not something you would use to make changes to the tune. Both sets of data have the same maximum, same minimum, and same average values. All cells have plenty of hits in them. The bottom data set has no data within +/- 5% of the average cell value. This means that the data does not correlate with the axis you plotted it against. The top case has 80%+ of the cell data within +/-5% of the average. This is valid data to modify your tune with.

Being able to quickly differentiate between the top and bottom data sets is crucial to improving a tune vs chasing your tail.



IMAG2775.jpg

[CCS86]

Crickets.

[Ben C]

It's a great idea, but won't make them money so bottom of the list.

[Keith@HPTuners]

It's a great idea, but won't make them money so bottom of the list.

That's not how we prioritize things here at HPT.

[CCS86]

It's a great idea, but won't make them money so bottom of the list.

That's not how we prioritize things here at HPT.

I would argue that proper analysis tools like this are a great selling point and could differentiate their product from competition.

If you are tuning just a slow sweep or WOT datalog, you don't need this.

But the way many people try to tune, by gathering a long log of mixed driving and plotting average trims against various parameters to match ECU tables, is a complete crap shoot without statistical analysis.

With the framework already in place, I don't think this would take an unreasonable amount of development time for the benefit it provides.

[scoob8000]

Makes perfect sense to me.

In my mind, the bigger your dataset, the more variables covered, the more accurate your data can be. That's why I was hoping for that merge functionality..

[CCS86]

Makes perfect sense to me.

In my mind, the bigger your dataset, the more variables covered, the more accurate your data can be. That's why I was hoping for that merge functionality..

For sure. But, the operative word is can. You still need to see which parameters your fueling correlates with. Confidence weighted coloring would tell you very quickly what to correct first: MAF transfer function, fuel flow / inferred rail pressure, pulse width, fuel load, MAF temp, etc.