All terms
Optimization
Policy Gradient
Reinforcement learning methods that adjust a policy directly using reward signals.
Definition
Policy gradient methods are a family of reinforcement learning algorithms that optimize a policy — the rule mapping situations to actions — by directly following the direction that increases expected reward. Rather than learning value estimates first, they nudge the policy to make high-reward actions more likely. They underpin training methods like PPO and GRPO (named recipes for this kind of reward-based tuning) that are widely used to align and fine-tune language models.