mdp

An MDP, or Markov decision process, is a mathematical framework for modeling decision making in stochastic environments. It represents a decision-maker interacting with an environment over a sequence of time steps, where the outcome depends on both the current state and the action chosen. The model is defined by a five-tuple (S, A, P, R, γ): S is a set of states, A is a set of actions, P(s'|s,a) is the probability of transitioning to state s' when taking action a in state s, R(s,a) is the immediate reward received, and γ ∈ [0,1) is a discount factor that prioritizes sooner rewards.

A policy π specifies the action to take in each state. It can be deterministic (π(s) ∈ A) or stochastic (π(a|s) provides a distribution over actions). The goal is to find an optimal policy π* that maximizes expected cumulative reward, typically defined as E[Σ_t γ^t R(S_t, A_t)]. Value functions V^π(s) and Q^π(s,a) are used to quantify expected returns under a policy. The Bellman equations relate these values: for a given policy, V^π(s) = Σ_a π(a|s) Σ_{s'} P(s'|s,a) [R(s,a) + γ V^π(s')], and the optimal Bellman equation V*(s) = max_a Σ_{s'} P(s'|s,a) [R(s,a) + γ V*(s')].

Algorithms for solving MDPs include dynamic programming methods such as value iteration and policy iteration when the model P and R are known; and model-free methods such as Q-learning, SARSA, and various policy-gradient or actor-critic methods when the model is unknown. In continuous state or action spaces, function approximation is often used. Extensions include partially observable MDPs (POMDPs), which handle hidden state information. MDPs underlie many applications in robotics, control, economics, operations research, and reinforcement learning.