AlgoPlus//machine learning / mdp
Read the theory

Markov Decision Process

States, actions, rewards — value iteration backs the reward up into a policy.

Iteration
Legend
High value / goal
Low value / pit
AI Tutor Workspace
In a nutshell
A Markov decision process (MDP) is the standard way to frame a decision-making problem: an agent in some state takes an action, lands in a new state, and collects a reward. Value iteration solves it by asking each state 'how good is it to be here?' — the answer is the best reward reachable through its neighbours, with future reward discounted a little. Repeating this backs value outward from the goal until the best action in every state is clear; those best actions together form the optimal policy.
Ready
Press play to begin the cinematic walkthrough.
Ask each square 'how good is it to stand here?' by looking at the best neighbour. Repeat, and worth ripples out from the goal until following the arrows is obviously optimal.
Key terms
Go deeper in the lesson
Read the full theory, intuition & complexity for Markov Decision Process.