Hacker News Reader

Comment by toxik | Hacker News Reader

Guys, the AI has perfect state information. While this is certainly impressive, it is not what you think it is.

This was one of the big advantages the AI had.

I could tell that the human pilot had to reorient himself after encounters, while the AI never needed to waste time grasping the situation and was able to take advantage immediately, every time.

In the same vein, as a former fighter pilot was commenting, the advantage goes to the pilot who can keep the plane right at the very edge of its performance envelope--which an AI can do more precisely than a human pilot can.

Classical dog fighting is all about energy management, something which an AI can outperform humans in considerably since it doesn’t uses intuition and experience but real time data.

With two identical airplanes, that's true. IIRC the AI also excelled in high-aspect head on shots which normal pilots tend to avoid, because for humans it's mostly about luck and less skill. When you pit two dissimilar aircraft against each other, it's up to the pilot to try to keep the fight outside of regimes where the enemy plane performs better.

A common comparison is between a one-circle and two circle fight, where the aircraft turn in opposite and same direction, respectively (and unintuitively!). Pure energy fighters like to keep going in a circle at the peak speed of its turn curve, when high-AoA capable planes want to cut into the enemy's flight path to get an early shot.

Another surprising example is a well-known HUD tape where a slow but fast-turning A-10 beats an F-16 in a training dogfight when the real-world solution would have been to not engage in a turning fight but hit the slower plane with a missile.

We know from John Boyd that getting into the opponent’s OODA (observe-orient-decide-act) loop is what wins these fights. If the AI has perfect state information, then it’s no wonder that it beat the human pilot who has to go through all four stages.

Exactly, if you could replicate that perfect state information from real world sensors, then this become much more impressive.

Or introduce noise into the simulated state to reproduce real world sensor capabilities.

I can see the scary part being what happens when you network a bunch of these. At least until the EWAR guys figure out how to ruin it's day.

how about this: the scary part is when the individual nodes are rigged to be able to operate autonomously, and are able to elect new leaders or form consensus with peers and select new mission objectives as required by the situation if the prior command hierarchy is destroyed or otherwise uncontactable for an extended period

Uh huh.

A) Easier said than done.

B) How do you protect against malicious node injection and compromise?

C) Where's the off switch?

If you haven't solved the problem amongst carbon based life forms, you really need to think twice before jamming it into silicon. Boring? Yes. Less likely to end up boomeranging into your face? Most definitely.

Reply on news.ycombinator.com