Yet a predictable system doesn't always lead to expected output.

I once heard a story about a student, probably graduate, who wrote a very complex genetic algorithm, complete with a physics engine, to build (out of blocks representing legs, arms, etc) a creature that could best walk to a finish line.

But what it actually did, after months of work and weeks of running, is stack all the blocks on top of each other and have the creature fall over the finish line, because it only measured how fast the creature reached the finish line, not whether the entire creature crossed it.

And don't tell me that isn't a creative (albeit rules-lawyering) solution to the problem.