A policy chooses an action from the information available to an agent. The environment responds with a new observation and reward, and training uses that experience to improve the policy. Training can use newly collected interactions or previously recorded experience; it need not occur inside a live customer application.
For example, a warehouse simulator can reward a routing policy for completing deliveries while penalizing collisions and delays. Practicing in a simulator or a digital twin avoids real collisions, but the policy still has to survive sim-to-real transfer to the actual warehouse. A move that seems slow immediately may prevent a later collision. Optimizing reward across a sequence makes the problem different from predicting an isolated labeled answer.
The reward definition matters because the learner can improve its measured score without achieving the intended goal. Rewarding distance traveled could encourage unnecessary driving. Reward hacking names such mismatches. Agentic reinforcement learning applies the same idea to tasks involving tools or other agent actions; connecting a tool alone does not train a policy.
The usual formal setup is a Markov decision process. Temporal difference learning updates value estimates from each step rather than waiting for a final outcome. Game-playing systems such as AlphaGo combined learned networks with Monte Carlo tree search to choose moves.
Sources
- OpenAI Spinning Up: Key concepts in RL — Defines policies, trajectories, environments, rewards and the expected-return objective.
Go deeper
- Hugging Face: Introduction to deep reinforcement learning course
Start with the agent-environment loop and work toward a trained game-playing agent.