Inverse reinforcement learning reverses the direction of ordinary reinforcement learning. In the usual setup you give an agent a reward function and it learns a policy that earns high reward. In inverse reinforcement learning you observe an expert’s behavior, assume it is roughly optimal, and try to recover the reward function the expert must have been pursuing. The motivation is that for many tasks, such as driving a car well, the reward is far harder to write down than it is to demonstrate.
The problem was framed by Andrew Ng and Stuart Russell in their paper “Algorithms for Inverse Reinforcement Learning,” presented at ICML 2000. They formalized it within Markov decision processes and gave three algorithms for extracting a reward function from observed optimal behavior. A central difficulty they identified is that the problem is ill-posed: many different reward functions can explain the same behavior, including trivial ones, so the algorithms need extra principles to pick a sensible answer.
Inverse reinforcement learning underpins a broad family of techniques for learning from demonstration and for aligning AI with human intentions, because it is a way to learn what people actually value rather than what someone hand-coded. Why business readers should care: it is the conceptual cousin of reinforcement learning from human feedback, which trains a reward model from human preferences, and it points at a general strategy for getting machines to do what is hard to specify but easy to show.