[4gotunameagain]

no serious, safety critical system uses RL (except tesla "autopilot" and we see how that went). Control theory algorithms can be validated to work within the desired envelope and produce a valid solution.

The big advantage of convexifying the problem, is that when it is convex you have a guarantee it can be solved in fixed time, a major requirement for real time systems

[PartiallyTyped]

I wasn't thinking of DeepRL, but more on the more classical side of things with approximators other than neural NNs; but what you describe makes sense.

[bo1024]

On that side, reinforcement learning bleeds over into control theory, so you're partly right.