What Is Reinforcement Learning
Reinforcement learning trains an agent to choose actions that maximize cumulative reward through trial-and-error interaction with an environment.
The learning problem
Reinforcement learning (RL) studies how an agent should act in an environment to maximize a scalar reward signal accumulated over time. Unlike supervised learning, the agent is never told the correct action. It receives only a reward that evaluates the consequences of what it did, and it must discover good behavior by experimenting.
At each time step the agent observes a state, selects an action, and receives a reward plus a new state. The loop repeats, generating a trajectory of experience. The agent's goal is to learn a policy — a rule mapping states to actions — that yields the largest expected return.
What makes it hard
- Feedback is evaluative, not instructive: reward says how good, not what was right.
- Consequences are delayed: an action's payoff may arrive many steps later (credit assignment).
- Data is not fixed: the agent's own choices shape the distribution of experience it sees.
- Exploration is required: the agent must try uncertain actions to discover better ones.
Where it applies
RL fits sequential decision problems: game playing, robotic control, resource scheduling, recommendation, and the control of complex physical plants. In fusion research, plasma-shape and coil-current control is a natural sequential-control problem; simulation environments can be used to train and evaluate controllers before any hardware exists. Kronos machines such as the Hyperion breeder are design and simulation studies, not built systems, so any control work today lives entirely in simulation.
The mathematical backbone of RL is the Markov decision process, which formalizes states, actions, rewards, and transitions.