RL vs planning

Both RL and planning aim to choose good actions, but they start from different places. Planning assumes you already have a model of the world—the rules of how states and rewards change—and computes a good plan by thinking ahead, like solving a maze on paper before walking it. RL instead learns from real experience, often with no model at all, discovering what works by acting in the world.

The line blurs in the middle. Model-based RL learns a model from experience and then plans inside it; classic planners like Monte Carlo tree search can be guided by learned values, as in AlphaGo. The honest contrast is: planning needs a known or learned model and pays in computation; learning from experience needs no model but pays in trial, error, and the data it must gather.