Thats some A.I/ Machine learning tech. Would need a lot of servers and a lot of engineers to make that work.
At best it could be a post game write up. Feed the replay to the AI coach, then receive performance review of what worked and what didn't.
Edit: if you're responding to the idea that Ace added, ignore my impassioned reiteration of my original idea. Showing what could have happened instead is indeed a level of complexity beyond what I proposed.
I don't think reporting descriptive statistics qualifies as machine learning. A lot could be done with simple frequency counts of attacks, whether they hit, miss, or were blocked.
You can calculate missed punish opportunities with simple frame data subtraction and character positioning.
To teach more optimal combos you just have to know character position and the first successful attack, as well as a list of possible combos which is quite finite.
I'm not a programmer so I'll certainly defer to others' expertise, but I believe the barrier to completing this task would be time and effort, not sophistication of data analytics and cutting edge programming.
I any case I agree that Capcom will never do this.
And it probably would be too much to calculate in real time, but I bet they could give *some* kind of feedback between rounds. Like "you're taking a lot damage from the air" "or you only blocked 15% of low attacks".
In conclusion, if their already existing AI can read inputs and beat your attacks, and tool_assisted can program a bot that does the same, this isn't pie on the sky stuff.
Even very simple feedback would give players more info than they'd otherwise have time to figure out in their own.