IMHO, stochastic optimal (SOC) control is a framework that is much more powerful than anything like current artificial intelligence, machine learning, deep learning, etc. and, in particular, a much better version of something like actual learning. The difference is like going from gnats to elephants, like a bicycle to the star ship Enterprise.

A good application of SOC looks not just "intelligent" but wise, prudent, brilliant, and prescient.

There have been claims that necessarily SOC is the most powerful version of learning there can be -- a bit much but ... maybe.

For computing architectures, the computing that SOC can soak up makes the computing needs of current deep learning look like counting on fingers in kindergarten.

The subject is no joke and goes back to E. Dynkin (student of Kolmogorov and Gel'fand), D. Bertsekas (used neural nets for approximations for some of the huge tables of data), R. Rockafellar (e.g., scenario aggregation), and, sure, R. Bellman. Some of the pure math is tricky, e.g., measurable selection.

I did my Ph.D. dissertation in SOC and there actually made the computing reasonable, e.g., wrote and ran the software -- ran in a few minutes, with my progress in algorithms down from an estimate of 64 years. For larger problems, can be back to taking over all of Amazon for 64 years. More algorithmic progress is possible, and, then, some specialized hardware should also help, sure, by factors of 10; right, we want lots of factors of 10.

If want to think ambitious computing, past Moore's law, past anything like current AI, with special purpose hardware, go for SOC.

SOC Applications 101. For your company, do financial planning for the next 50 years, 60 months. So, set up a spreadsheet with one column for each month, one column for the current state of the company and then 60 more columns. Goal is, say, to maximize the expected value of something, maybe the value of the company, in the last column, right, with control over the probability of going broke, if a bank always be able to meet reserve requirements and pass stress tests; if a property-casuality insurance company, stay in business after hurricane Florence, etc.

For each variable of interest, have a row.

In the cells, put in the usual expressions in terms of the values of cells in earlier columns.

Also have some cells with random numbers -- the stochastic part.

Also have some cells empty for the business decisions, the control part.

This is the 101 version; the 201 version has more detail!

Can't get the solution with just the usual spreadsheet recalc because the best solution is optimal and varies through the 60 months as more information is gathered -- and that is the core of the need for more in algorithms and taking over all of Amazon for the computing.

Really, the work comes too close to looking at all possible business state scenarios over the 60 months yet still is astronomically faster than direct or naive ways to do this.

Or, a little like football, don't call the play on 2nd down until see the results of the play on 1st down, BUT the play called on 1st down was optimal considering the options for the later downs and plays.

The optimality achieved is strict, the best possible, not merely heuristic: No means of making the decisions using only information available when the decisions are made can do better (proved in my dissertation).

Intel, AMD, Qualcomm, DARPA, NSF, Microsoft, etc., think SOC!!! You just heard it here -- I might not bother to tell you again.