DARPA AlphaDogfight Finals: AI agents compete to face a human F-16 pilot

(3) IDK what their vision model is in the simulation, but you are correct – keeping sight is vital in dogfighting. The F-16 is formidably tiny. It's practically invisible nose-on. (2) Air Combat Maneuvering ("ACM" or "dogfighting") training typically begins each engagement a mile or so abeam. I'd say that more than 95% of my 1v1 exercises began this way. So no radar lock, visual only. (1) See previous answer: the engagements begin with no radar lock and the turning characteristics of modern fighters rapidly compress the fight "inside a phone book." So there's a constant tradeoff of energy and geometry. Events often occur inside the missile-arming ranges. Radar and/or IR locks are fleeting.

Most radar/missile training is done in 2v2 (or sometimes mvn) engagements. These are usually commenced head-on, with "fight's on" signal given at the first pass.

One can argue about how representative or realistic any of this training is. But it does make you a better aviator. And fighter pilots love it.

Also, my experience is from decades ago. Take it FWIW.