Even if you had a perfect model of your amplifier, it would need to be fed with information which will only be available after errors have already been committed.
Take the fact that each temperature gradient over every resistor will lead to non-linearities because of the thermocouple effect, potentially creating distortion products. So, you would need to monitor this for all relevant resistors, but these measurements will have a certain time delay. This will lead to errors, since the predistortion model of the amplifier will no longer be able to accurately predict it's future behaviour.
Therefore, no pre-distortion system will ever be able to see into the future; it cannot even predict it with perfect resolution.
vac
The state of the art in neural nets just isn't capable of anything close to what this thread is about. But I gave you a longer answer privately in the WC thread. 🙂
I wonder how the predistortion mechanism would be implemented ?
Could that be analog ?
Or will it be an A/D conversion, table lookup, D/A conversion.
Could that be analog ?
Or will it be an A/D conversion, table lookup, D/A conversion.
By seeing into the future, I mean that the input to the ANN can include past and future samples relative to the output, so the correction can actually be derived from future samples. (The overall output is delayed by some arbitrary amount relative to the input).
It's true that the training is derived from errors, but unlike ordinary negative feedback the 'error term' can be anything we choose. If, for example, the ANN training finds that responding in a certain way to a step input produces some instability that always affects future output from the amplifier, it can make do with less perfect correction for sample N in order to get a better average error for samples N to N+10, say. Neural networks (or similar structures) can theoretically achieve this sort of thing - but we wouldn't know they were doing it, necessarily.
As I say, I don't know there's anything to worry about with ordinary feedback, but many people around here seem to think that amplifiers sound different even though they all measure blamelessly. The finger of suspicion seems to point to problems with feedback on transients. Using predistortion, the nature of the correction could be tailored to, effectively, change depending on the characteristics of the past, present and future input, which I can't see being possible with ordinary feedback.
It's all just theoretical, I know, but I would be thinking that it would have to be ADC (or CD/WAV) -> CPU -> DAC, or maybe direct digital to a Class D type output.
- Status
- Not open for further replies.