Reinforcement learning

Objectives

– Objective 1: be able to model a situation as a reinforcement learning problem (choice of reward function, state space, action space, etc.)

– Objective 2: be able to learn a policy from data and evaluate it

– Objective 3: understand the limits of reinforcement learning modeling

Course outline

– Bandit models

– Exploration-exploitation dilemma

– Bandit algorithms

– Regret

– Markov decision process (MDP)

– Bellman equation

– State-value function

– Algorithms for MDP

Prerequisites

Deep learning and Markov chains