Is There A Need For Stochastic Inputs To Mimic Real-World Biology And Environment?

Question

I'm kind of new to machine learning/AI, but I was wondering if using thresholds/fuzzy logic-like functions and even networks of dependent, stochastic variables that change over time (LTL maybe?), would be ample enough to emulate natural processes like emotions, hunger, maybe even pain.

My dilemma is whether creating a basic library to do this for the developer community is worth it if everything can be modeled more-or-less mathematically deterministic, even if the formulas are really complicated (see research like: https://engineering.stanford.edu/news/virtual-cell-would-bring-benefits-computer-simulation-biology).

My initial reasoning was biological processes are connected to psychological functionality (e.g., being hungry might make someone irritable, but that irritability may wear-off, which triggers different paths of thought but not others). But these are so inter-dependent that it may be random or it is essentially PRNG, in order to properly simulate the mood fluctuations and biological processes computers don't have but humans do have.

Would we be better-off waiting for these complex physical/neurological models to come out?

Answer 1

Quite an interesting question, this sort of computational representation of biological systems are at the forefront of what we are trying to accomplish algorithmically.

This can be tackled via a variety of coordinate frames and problem formulations and could be thought of quite philosophically, as it pokes at some deep questions such as free-will and consciousness more generally.

For example, we could take the assumption that based on everything we know physically, the universe is completely deterministic. In this case, implementing deterministic functions for these variables makes quite a bit of sense and seems like the obvious choice. However, we can easily imagine that for at least some of this dynamic variables a stochastic representation would be both simpler(computationally) as well as giving better approximations to what we see in the real-world.

So it really is a question of how much value you see from implementing a system like this, and which representation you think would be most effective.