Supervised learning, the most commonly discussed form of machine learning, trains a model using examples that already come pre-labelled with the correct answer, letting the model directly compare its own prediction against a known right answer at every single training example. Reinforcement learning works under a genuinely harder constraint: an agent takes actions within some environment and receives only a reward signal afterward, indicating roughly how good or bad the outcome was, with no labelled examples telling it directly what the correct action actually would have been at any given point — it has to work out for itself, purely from that sparse reward feedback, which of its own past actions actually contributed to earning it.
No labelled "correct action" exists for reinforcement learning to check against
A supervised learning model training to recognise images, for instance, is given a large set of images each already labelled with the correct category, and it directly adjusts itself to reduce the specific gap between its own prediction and that known correct answer for every training example it sees. A reinforcement learning agent has no equivalent labelled "correct action" to check itself against at any given point — it only receives a reward signal describing how good or bad an outcome turned out to be, sometimes only after a long, extended sequence of many individual actions, leaving it to work out on its own, purely through trial and error across repeated attempts, which of its own earlier individual actions genuinely contributed to that eventual reward and which didn't matter much at all.
The credit assignment problem: figuring out which action actually earned the reward
This core challenge, working out which of a long sequence of individual past actions genuinely deserves credit or blame for an eventual reward that only arrives later, is known as the credit assignment problem, and it's one of the central technical difficulties reinforcement learning specifically has to solve that supervised learning, with its example-by-example labelled feedback, simply doesn't face in the same way. Reinforcement learning algorithms generally address this by repeatedly trying different sequences of actions across many attempts and gradually building up an estimate of how much each individual action, in each specific situation, genuinely tends to contribute to the reward eventually received, refining that estimate over a large number of repeated trials rather than being told the correct answer directly from a single labelled example the way supervised learning is.
What we're still unsure about
The basic distinction between supervised learning's labelled-example training and reinforcement learning's sparse reward-based training, along with the credit assignment problem, are well-established, foundational concepts in machine learning research. What remains a genuinely active area of ongoing research is designing reinforcement learning algorithms that can learn efficiently even when rewards are extremely sparse or delayed over very long sequences of actions, since the basic trial-and-error approach can require an enormous number of attempts to reliably work out which actions actually mattered when a meaningful reward only ever arrives after a very long sequence — researchers continue actively developing new techniques specifically aimed at this sample-efficiency problem, without a single universally solved answer that works equally well across every kind of reinforcement learning task.
This sits inside Reinforcement Learning, one of eight topics in AI/ML, one of seven domains in Computer Science, one of seventeen subjects the app can quiz you on.