I found myself wondering about this based on some recent experiences with an interesting project. Since this is ongoing work whose direction I am still pondering, let me just say that this project involves some fairly complex dynamics of the sort that can be described by a nonlinear partial differential equation. It is complex, but well modeled – in the sense that there is an established physical theory underlying this process.
My involvement in this has been from the diagnostics and control side. The question that got me started was this – is it possible to use machine learning methods to acquire good low-dimensional models of the process, so that one can use that in real time loops as a replacement for a more computationally intensive PDE solver in the loop? I have several hypotheses for my specific problem that I will be working through during the coming months.
But, in general, if you know the structure of the dynamics and if the problem ‘falls’ with just a few more processors then doesn’t ML become somewhat redundant. Isn’t this nearly always the case in engineering? Then why waste time with these methods?
So, I have been pondering (in the background) this for several days now.
As it happens, this is a complete non-issue for my main research thread – robotics. One reason for this is that goals are ill-specified (unlike in the above problem, where the goal is simple – predict the current configuration, or stabilize configuration X), we need complex global behaviors and the system is subject to large perturbations by the environment. So, just formulating what it means to do task X requires some care – and learning methods can help push many of these decisions to their proper place inside the robot’s head. Moreover, this is just an indication of the larger point – that I am using robots as a means to asking and answering scientific questions about learning and cognition as they relate to robust autonomous behavior, so no further justification is necessary – and even if it were, the ability to achieve this sort of robust autonomy with bounded resources is desirable in practice.
This is probably not news to anyone who has thought about ML seriously. However, occasionally, I do run into people who like to brag about the “complexity” of the dynamics involved in their problem, and question whether ML can take it on. Here, I am saying that – more often than not – the exotic system may actually be easier to deal with (given proper and competent thought) than some seemingly simpler but much more ill-posed problems that we might never suspect of requiring serious research effort.