Mathematical foundations of RL: Markov decision processes, Bellman equations, policy gradients, and temporal difference methods.
9 concepts
A Markov Decision Process (MDP) models decision-making in situations where outcomes are partly random and partly under the control of a decision maker.
Bellman equations express how the value of a state or action equals immediate reward plus discounted value of what follows.
Temporal Difference (TD) Learning updates value estimates by bootstrapping from the next state's current estimate, enabling fast, online learning.
Proximal Policy Optimization (PPO) stabilizes policy gradient learning by preventing each update from moving the policy too far from the previous one.
Value function approximation replaces a huge table of values with a small set of parameters that can generalize across similar states.
The exploration–exploitation tradeoff is the tension between trying new actions to learn (exploration) and using the best-known action to earn rewards now (exploitation).
RLHF turns human preferences between two model outputs into training signals using a probabilistic model of choice.
The discounted return G_t sums all future rewards but down-weights distant rewards by powers of a discount factor γ.