Reinforcement Learning - Markov Decision Process

Introduction

• In Supervised learning we are given x and y and we should find the f(x)using function approximation techniques.
• In Unsupervised learning we are give x and we should using clustering methods to find the f(x)

Reinforcement Learning

In reinforcement learning we are given x and z we should find y and f(x)

Markov Decision Process

• States- Each step or states
• Actions - Actions that can be taken in a particular state
• Models - Transition steps in which it changes from one state to another state using particular action
• Rewards - Rewards you get when you perform some action or when you reach a state or when you reach a state from another state using an action

Properties:

• Only Present state matter
• Actions/States are stationery

Policy

Given a particular state a policy will suggest you a particular action to be taken

Optimal Policy

It is a policy which will help in optimize in receiving a maximum rewards.

Hence in reinforcement learning x refers to state, z refers to rewards y refers to the actions and f(x) refers to the optimal policy

Temporal Credit Assignment

Identifying the action of which made us to get the rewards or for the given state we are in what was sequence of actions that was taken.

Carefully choosing the rewards will determine the end results.

Sequence of Rewards

• Infinite Horizons

  We have ample time or no of steps to reach destination. Hence being stationery in fine.
  

• Finite Horizons

  Depending on the time or no of steps our actions differs
  Hence policy can be defined as the factors of both states and times
  

• Utility of Sequences

  Rewards we get via a sequence of states