What's interesting is that our modern computational platforms have capabilities that far exceed most of our use cases. And this is a good demonstration of why.
What's interesting is that our modern computational platforms have capabilities that far exceed most of our use cases. And this is a good demonstration of why.
I have an example from my daily work. We were hired to update a company's model from 6 years ago. Their model is an Excel VBA model that we're rebuilding in Python with an increase resolution of detail and accuracy. However, for our "more accurate" model to be accepted, we have to also prove the following:
- We have to show that our model's outputs are the same as the VBA's model for the same input/starting condition (makes sense at a higher level, but at a higher resolution this means you have to spend more time looking at why a specific value is in a specific way and can be different from the lower resolution model)
- Train the new users on the new model (knowledge transfer)
- Ability to update and modify the new model wherever they need to change it or understand it
- Understand the limitations and the constraints of the new model
This basically boils down to how much time is needed in training the new model, how much money and time is needed to invest in building a new model, and is it really worth it to build a new model when the old model has worked good enough for now (and you understand the limitations of said model and is within your needs)? There's only so much time in the day and you obviously have a ton of other tasks that also needs your attention, and getting these newer models verified and accepted by the board is great but it's also additional meetings and time that you need to setup with them and get everything OK-ed... doubly hard when everyone else is also super busy.
I'd say modern computational platforms are better when it comes to raw power, but the logic and just resources required to build, verify, accept, and train (human resources-wise) the new model might just not be fully worth it.
Remember, there are also other industries other than tech that uses computer hardware to get jobs done.
I think that this guy is using the analog computer because he's an enthusiast. That's fine. It's not better by any means other than making an employee happy.
Estimating accumulated error in floating-point calculations is an extremely difficult science. On the other hand, a simple circuit that solves an ODE in an analog computer can be analyzed rather easily.
If you devote enough memory, you can be much more precise with continuous values than an analog computer could ever hope to be. Frankly it wouldn't take much memory at all by modern standards though.
You may be thinking of the "just use enough bits so you're sure the accumulated error isn't significant" approach, which is essentially what modern floating-point attempts to do.
There are still some pitfalls though, around doing an analysis that proves the bounds of said error in a given algorithm and data set.
You might be interested in the concepts used in interval arithmetic [1] and specifically the work of Gustafson in designing some models for using those in scientific computing hardware.