[andyjohnson0]

The title seems misleading to me. The AI isn't finding bugs by somehow examining the game's source code, it's trying random gameplay and exploiting any advantages that emerge. That it's finding previously unknown bugs seems to be almost entirely down to trying things that human players wouldn't think to do.

[tantalor]

You confuse bug (unintended behavior) with its cause (bad code).

[BatFastard]

We called them "Unintended features", and they were usually quite popular with users.

[montyf]

Exactly what I was thinking. It may be a bug, but the AI treats it as another legitimate game rule. I wonder if there are any techniques for it to be able to tell the difference... for example, if it can quantitatively demonstrate that the conditions for a rule are very rare/unlikely.

[pvg]

The AI isn't finding bugs by somehow examining the game's source code

The title doesn't say that.

[PeterisP]

It kind of implies that - "AI Cheats at Old Atari Games by Finding Unknown Bugs" would be an accurate title, but the extended "AI Cheats at Old Atari Games by Finding Unknown Bugs in the Code" tells that it's actually finding something in the code, as opposed to simply unexpected/emergent behavior.

[pvg]

It's mildly ambiguous (like most things) but it's not misleading or inaccurate. It finds bugs. Which are in the code. It doesn't say what kind of code and it certainly doesn't say it finds them by looking at the source code.

[corobo]

I read it and presume it's meant to be read as the AI finding "unknown [bugs in the code]" as opposed to "unknown [bugs] in the code"

[soared]

Did you read the article?

>It’s important to note, though, that the agent is not approaching this problem in the same way that a human would. It’s not actively looking for exploits in the game with some Matrix-like computer-vision.

[andyjohnson0]

I did. "looking for exploits in the game with some Matrix-like computer-vision." is a fairly meaningless phrase.

[bcoates]

Seems like a perfectly obvious metaphor? It's just providing GA-generated input, not doing tooled exploration like afl-fuzz or a symbolic execution analyzer.

It literally-and-metaphorically can't see behind the UI to the internal state of the game.

[yorwba]

I haven't read the article yet, and I was not mislead by the title. I guess it helps to be familiar with the way reinforcement learning agents are hooked up to a simulation environment.