AI controlled ECU

[Zynerji]

In a super-saver formula like the 2026 rules seem to indicate, will we see full AI control of these multi-unit contributions and resources?

I can see a rainbow-table like GPT that simply responds to the torque demand of the pedal and the recharge of the ERS in the most efficient manner possible.

It almost seems inevitable that it then starts making recommendations based upon discovered limitations that reduce efficiency.

Are we there for 2026, or are we looking at the 2032 formula for that?

[Farnborough]

How would AI define efficiency in that scenario ?

[dialtone]

rainbow table doesn't seem AI, just a simple lookup. What would you gain with AI here? Seems a simpler calculation than what you'd have with AI and in order to run AI you'd need specialized hardware to do it fast enough anyway.

[Zynerji]

How would AI define efficiency in that scenario ?

It would monitor power in/power out. Fuel consumption etc.

[Zynerji]

rainbow table doesn't seem AI, just a simple lookup. What would you gain with AI here? Seems a simpler calculation than what you'd have with AI and in order to run AI you'd need specialized hardware to do it fast enough anyway.

Rainbow table in the sense that it would start with all of the precomputed combinations of settings, and then monitor patterns and such. Even to the point that it could become anticipatory of the current driver that it is learning from.

And APU processors with integrated tensor cores isn't hard. Especially if you were to comission a custom part from AMD.

[Farnborough]

How would AI define efficiency in that scenario ?

It would monitor power in/power out. Fuel consumption etc.

Current and simple systems enact that anyway, as described there. Nothing needed on that front.

How an it judge the most effective strategy against race "need" to have any meaningful control ?

Thats if I understand what the application of AI is supposed to do in the scenario you've set out.

[Farnborough]

rainbow table doesn't seem AI, just a simple lookup. What would you gain with AI here? Seems a simpler calculation than what you'd have with AI and in order to run AI you'd need specialized hardware to do it fast enough anyway.

Rainbow table in the sense that it would start with all of the precomputed combinations of settings, and then monitor patterns and such. Even to the point that it could become anticipatory of the current driver that it is learning from.

And APU processors with integrated tensor cores isn't hard. Especially if you were to comission a custom part from AMD.

With the existence of "pre conceived tables" is that then AI....or as noted above....just a lookup facility ?

[Greg Locock]

Most systems that are touted as AI driven use a pre-trained static model (often a neural network). There's no new 'thought' or learning going on, if you give it the same inputs you'll get the same outputs.

One exception is adaptive controls in cars which is why your gear changes get smoother with time, and also your engine uses the knock sensor and cat sensors to refine its tune, dynamically. But they aren't AI really.

[dialtone]

Most systems that are touted as AI driven use a pre-trained static model (often a neural network). There's no new 'thought' or learning going on, if you give it the same inputs you'll get the same outputs.

One exception is adaptive controls in cars which is why your gear changes get smoother with time, and also your engine uses the knock sensor and cat sensors to refine its tune, dynamically. But they aren't AI really.

AI is always a static model that you trained and deploy. There isn't really any advantage in re-training your model after each input, and for complex models with lots of inference it's not even possible. Using a neural network is helpful when the combination of inputs is too large to be practical to maintain or when the input isn't very easy to model via features or isn't structured anyway (like an image or free text).

If the optimization problem gets hard enough mathematically then machine learning can "do the math for you" and find that optimal point. I'm not saying it wouldn't be useful to have more intelligence in the ECU, I just wonder what it would be because everything feels very easy to model right now, and if there's an easier way to model the problem I would normally prefer that since it would be more reliable and deterministic while AI can have a more probabilistic result depending on what it is you are using.

[Zynerji]

I guess I was just picturing a fine-tuning type algorithm that would operate between the ECU control pulses, if that makes sense.

My thought was that this "tweening" would lead to better efficiency and/or performance gains. Even if only .5% but worth as software carries zero weight penalty.