Markov Decision Processes

Also known as: MDP

framework · mathematics · formal-scientific

Sequential decision framework with states, actions, transition probabilities, and rewards.

Markov Decision Processes (MDPs) are the mathematical framework for modeling sequential decision-making in stochastic environments where outcomes are partly random and partly under the control of a decision maker. An MDP is specified by: a set of states S; a set of actions A; transition probabilities P(s'|s,a) — probability of next state s' given current state s and action a; reward function R(s,a) — immediate reward for taking action a in state s; discount factor γ in [0,1] for valuing future vs immediate rewards. The Markov property — that transition probabilities depend only on current state and action, not on history — is the central simplifying assumption. Solving an MDP means finding an optimal policy — a mapping from states to actions — that maximizes expected discounted cumulative reward. The framework was substantially developed by Richard Bellman in his 1957 Dynamic Programming, introducing the Bellman equation and the principle of optimality, and elaborated by Ronald Howard's 1960 Dynamic Programming and Markov Processes. MDPs are foundational to operations research, control theory, and reinforcement learning (where the agent typically doesn't know transition probabilities or rewards in advance and must learn them through interaction). Modern reinforcement learning — Sutton and Barto's foundational Reinforcement Learning: An Introduction (1998, 2018), Q-learning (Watkins 1989), policy gradient methods, deep reinforcement learning (DeepMind's DQN 2013, AlphaGo 2016) — substantially builds on MDP foundations. Variants include Partially Observable MDPs (POMDPs, where the agent doesn't directly observe state), continuous-state and continuous-action MDPs, and risk-sensitive MDPs.

Originators

Richard Bellman (foundational 1957 Dynamic Programming); Ronald Howard (1960 Dynamic Programming and Markov Processes); subsequent foundational reinforcement-learning figures including Christopher Watkins, Richard Sutton, Andrew Barto high

Year / Decade

1957 (Bellman); 1960 (Howard); 1989 (Watkins Q-learning); 1998-2018 (Sutton-Barto) high

Primary sources

Bellman, R. (1957). Dynamic Programming, Howard, R.A. (1960). Dynamic Programming and Markov Processes, Sutton, R.S. & Barto, A.G. (1998, 2nd ed. 2018). Reinforcement Learning: An Introduction, Puterman, M.L. (1994). Markov Decision Processes high

Core components

Primary use case

Foundation of modern reinforcement learning; operations research and stochastic optimization; control theory; foundation for substantial AI work including game-playing agents (AlphaGo, AlphaZero), robotics, autonomous vehicles, recommendation systems; medical decision-making (treatment policies under uncertainty); resource management and scheduling; foundation for many modern AI products and research directions.

Common criticisms

Lineage

Child of
Decision Theory
Siblings
Decision Theory, Bayesian Inference
Derived from
Decision Theory