Environment Wrappers — ELI5

Imagine you have a pair of binoculars. By themselves they show you a normal view. But you can snap on different lenses: a zoom lens, a colour filter, or a night-vision attachment. The world outside does not change — only what you see through the binoculars changes.

Environment wrappers in reinforcement learning work exactly like those snap-on lenses. The game (called the environment) stays the same, but a wrapper sits between the game and the learning program, tweaking what the program sees or how its actions are interpreted.

One wrapper might shrink a big colourful image down to a small grey one, so the program processes it faster. Another might clip the score so it never goes above one or below negative one, which helps the program learn more steadily. A third might stop the game early if it has been running too long.

The beautiful part is that wrappers stack. You can snap on a resize lens, then a colour filter, then a score limiter — all at once. And because each wrapper only changes one thing, you can mix and match them for different experiments without touching the game code at all.

This is why wrappers are everywhere in reinforcement learning. Researchers rarely use a raw game. They wrap it with a stack of helpers that massage the inputs and outputs into a shape their learning algorithm handles best.

The one thing to remember: Environment wrappers are snap-on lenses that change what a learning program sees and how its actions are handled, without modifying the game itself.