Reinforcement Learning

Reinforcement learning is a training approach in which an agent learns by acting in an environment and receiving rewards, rather than by copying labelled examples. At each step the agent observes a state, chooses an action under its policy, and gets a reward signal; over many episodes it shifts the policy toward actions that raise expected long-term return. The central difficulty is credit assignment, since a reward may arrive long after the move that earned it, and the agent must balance exploring new actions against exploiting what already works. DeepMind's AlphaGo and AlphaZero learned board games this way, and the method drives robot control and data-centre cooling. In language models it appears as a post-training stage: RLHF trains a reward model on human preference comparisons and then optimises the policy against it, RLAIF replaces human raters with a model, and DPO reaches a similar result without a separate reward model. Reward hacking is the standing risk. An agent optimises the reward it is given, not the intent behind it, so a poorly specified reward produces behaviour that scores well and misses the point, such as answers that sound agreeable rather than accurate.