Simply put, because you often want to change the state of something without breaking all the references to it.

Wild off the top of my head example: you’re simulating a football game. Everything is represented by objects which hold references to other objects that are relevant. The ball object is held by player object W, player object X is in collision with and holds a reference to player object Y, player Z is forming a plan to pass to player object X (and that plan object holds a reference to player object X) and so on.

You want to be able to change the state of the ball object (its position say) without creating a new object, because that would invalidate how every other existing object relates to the ball.

“we heavily biased the network against trains and now it’s just saying the optimal car consists of several metal struts connecting just two thinned out wheels that the driver sits on top of and propels themselves using pedals. It was busy redesigning intersections to have clear safe lanes for these bi-cycle ‘cars’ with plenty of trees / room for pedestrians when we pulled the plug…”