Critiquing the AI ASW

[SeaQueen]

Originally Posted by LoBlo

Which begs the question... what *should* a half-way decent response to a hostile sub be? Sure they will shoot ASROCs and helos at the sub, but it seems like course changes and defensive sprints are lacking.

This is actually a problem that comes up with real world ASW modeling problems as well. CLEARLY a prudent commander would order the force to change course if he had enough warning, but to what course? The answer is, as usual, "it depends," and for programming an AI, that gets hard. There isn't any single assumption you can make.

I don't know if you guys remember me harping about the limiting lines of approach? I think I put an example of it in the NATO EX WAR Exercise and RIMPAC CSG scenarios. It's just kinematics and geometry. The idea is that if the submarine is slow relative to the group of warships, then there's only a limited space in the ocean, from which a submarine could approach the warships and end up within it's weapons launch range.

From the perspective of a group of warships trying to defend themselves against a submarine then, the goal would be to turn so that the limiting lines of approach do not intersect the area of uncertainty surrounding their initial sub contact.

This is where the decision making gets fuzzy. What assumptions do you then make about how fast the enemy submarine might be going? What assumptions do you make about the enemy weapons? Is the intelligence informing your assumptions correct? If it's inaccurate, how inaccurate is it? If you don't even know what type of submarine you're facing, what assumption do you make? Well... if it was me I'd assume it was the fastest submarine out there which would be a nuclear boat. Okay... but a nuclear boat might not be constrained by the limiting lines of approach. They can easily be faster than a lot of ships. It gets very difficult to program an AI that would accurately reproduce the kinds of thinking that would go on in all situations.

The whole thing is scenario dependent. What adversary are you facing? How much do you know about that adversary? Can he go fast and silent or does he cavitate if he goes much faster than 7 knots? That's the sort of thing which only ever really gets captured in multiplayer games. There's other things too. Let's suppose you do make the correct assumptions about the nature of the adversary or at least guess a good direction to turn at random. Now you've got to decide how quickly to turn. If you detect him far out, then maybe you can evade with a gentle turn and laugh the whole thing off as your helos or an orbiting P-3 prosecute him. If you detect him close in, it might not matter how quickly you turn because he's almost in torpedo launch range anyhow.

This is the stuff that makes wargaming fun if you ask me. Unfortunately, it's also why AIs will always be boring. In order to capture the uncertainty in the decision making, they either have to be painfully stupid in the sense that they make excessively conservative assumptions about course changes and what not, or else they end up being so random that you're left thinking, "what the heck is that thing doing?"