Scribd
Upload
8 views

Lecture Week12

The document provides an overview of reinforcement learning, including elements, use cases, agents and states, components of an RL agent, exploration vs exploitation, challenges, and multi-agent reinforcement learning. It also covers differences between supervised, unsupervised and reinforcement learning and discusses policy, value functions, and models in RL agents.

Uploaded by

Reedus
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
8 views

Lecture Week12

The document provides an overview of reinforcement learning, including elements, use cases, agents and states, components of an RL agent, exploration vs exploitation, challenges, and multi-agent reinforcement learning. It also covers differences between supervised, unsupervised and reinforcement learning and discusses policy, value functions, and models in RL agents.

Uploaded by

Reedus
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 37

Reinforcement Learning:

Introduction
Lecture 12
Dr. Syed Maaz Shahid
20th May,2024
Contents
• Overview
• Elements of Reinforcement Learning
• Use Cases
• Agents and States
• Components of an RL Agent
• RL Agent Taxonomy
• Exploration and Exploitation
• Major Challenges of RL
• Multi-Agent Reinforcement Learning
Assignment
• Find a paper that uses HMM to solve a problem in your relevant
field.

• Make a report and submit by 31 May, 20204.


Machine Learning
Reinforcement Learning
• In supervised machine learning, you focus on predicting what you
don’t know→ learning to predict.

• Reinforcement learning (RL) is a framework for learning-based


decision-making→ learning to act.

• Reinforcement learning uses rewards and punishments as signals for


positive and negative behavior.
Reinforcement Learning: Elements
• Agent/ Policy: The program you train for specific task/Method to map
agent’s state to actions.
• Environment/State: The world, real or virtual, in which the agent
performs actions.
• Action. A move made by the agent, which causes a status change in
the environment.
• Rewards. The evaluation of an action (positive or negative)/Feedback
from the environment..
Reinforcement Learning: Elements
Reinforcement Learning: Use Case
• PacMan game
• Agent → PacMan
• Environment: Grid world
• Reward for eating food & Punishment if gets killed by the ghost
• State→ Location
• Cumulative reward→ winning the game
Reinforcement Learning: Use Case
• Battery Management for Retired Electric Vehicle Batteries

Doan, Nhat Quang, et al. "Deep Reinforcement Learning-Based Battery Management Algorithm for Retired Electric Vehicle Batteries with a Heterogeneous State of
Health in BESSs." Energies 17.1 (2023): 79.
Reinforcement Learning: Use Case
• Deep Q-learning playing Atari

• AlphaGo - The Movie


Differences: Supervised, Unsupervised, and
Reinforcement Learning
• Static vs. Dynamic

• No explicit right answer→ agent learns by trial and error

• RL requires exploration

• RL is a multiple-decision process→ decision-making chain through


the time in RL
Reinforcement Learning: Agent & Environment
• At each step t the agent
• Executes action 𝐴𝑡
• Receives observation 𝑂𝑡
• Receives scalar reward 𝑅𝑡
• The environment:
• Receives action 𝐴𝑡
• Emits observation 𝑂𝑡+1
• Emits scalar reward 𝑅𝑡+1
Reinforcement Learning: History & State
• The history 𝐻𝑡 is the sequence of observations, actions, rewards

• State is the information used to determine what happens next.

• Formally, state is a function of the history:


𝑆𝑡 = 𝑓(𝐻𝑡 )
Reinforcement Learning: Environment State
• The environment state 𝑆𝑡𝑒 is the
environment’s private representation.

• Usually, 𝑆𝑡𝑒 is not visible to the agent.

• Even if 𝑆𝑡𝑒 is visible to agent, it may contain


irrelevant information
Reinforcement Learning: Agent State
• The agent state 𝑆𝑡𝑎 is the agent’s internal
representation→ uses to pick the next action.
•
• It can be any function of history:
𝑆𝑡𝑎 = 𝑓(𝐻𝑡 )
Reinforcement Learning: Information State
• An information state (a.k.a. Markov state) contains all useful
information from the history.
• Markov chain is to assume that 𝑋𝑘 captures all the relevant information
for predicting the future

• A state St is Markov if and only if


𝑃 𝑆𝑡+1 𝑆𝑡 = 𝑃[𝑆𝑡+1 |𝑆1 , … , 𝑆𝑡 ]

• The future is independent of the past given the present


• Once the state is known, the history may be thrown away
Reinforcement Learning: Fully Observable
Environments
• Agent directly observes environment state
Ot = 𝑆𝑡𝑎 = 𝑆𝑡𝑒

• This is a Markov decision process (MDP).


Reinforcement Learning: Partially Observable
Environments
• Agent indirectly observes environment
𝑆𝑡𝑎 ≠ 𝑆𝑡𝑒

→A poker playing agent only observes public cards.

• This is a partially observable Markov decision process (POMDP).

• Agent must construct its own state representation 𝑆𝑡𝑎 = 𝐻𝑡


Components of an RL Agent
• Policy: agent’s behavior function

• Value function: how good is each state and/or action

• Model: agent’s representation of the environment


Policy
• A policy is the agent’s behavior.

• It is a map from state to action,

• Deterministic policy: 𝑎 = 𝜋(𝑠)→same action give a state

• Stochastic policy: 𝜋(𝑎|𝑠) = 𝑃[𝐴𝑡 = 𝑎|𝑆𝑡 = 𝑠]→ choose an action


randomly based on the probability distribution
Value Function
• Value function 𝑣𝜋 (𝑠) is a prediction of future reward

• Used to evaluate the goodness/badness of states

• By following a policy 𝜋, the value function is defined as


𝑣𝜋 𝑠 = 𝐸[𝑅𝑡 + 𝛾𝑅𝑡+1 + 𝛾 2 𝑅𝑡+2 + ⋯ |𝑆𝑡 = 𝑠]

• 0 ≤ 𝛾 ≤ 1: discount rate
• 𝛾 close to 1: rewards further in the future count more→ agent is farsighted
Model
• The model describes the environment by a distribution over
rewards and state transitions.

𝑃𝑠𝑎′ = 𝑃 𝑆𝑡+1 = 𝑠 ′ 𝑆𝑡 = 𝑠, 𝐴𝑡 = 𝑎

𝑅𝑠𝑎′ = 𝐸[𝑅𝑡+1 |𝑆𝑡 = 𝑠, 𝐴𝑡 = 𝑎]


RL Agent Taxonomy
RL Agent Taxonomy
• DQN→ value-based method

• A2C→ policy-based method (actor) & value-based method (critic)

• Policy Optimization (PPO)


Taxonomy of Reinforcement Learning
algorithms

OpenAI spinning up
Reinforcement Learning algorithms
• Model-based algorithms use a model of the environment.
• Model is used to predict future states and rewards.
• The model is either given (e.g. a chessboard) or learned.

• Model-free algorithms directly learn how to act for the states


encountered during training
• Which state-action pairs yield good rewards (Q-Learning).
Reinforcement Learning: Fundamental
Problems
• Learning:
• The environment is initially unknown → which states are good or what the
actions do.
• The agent interacts with the environment
• The agent improves its policy

• Planning:
• A model of the environment is known→ Markov decision problem
• The agent performs computations with its model
• The agent improves its policy
Exploration and Exploitation
• Reinforcement learning is like trial-and-error learning.
• The agent should discover a good policy from its experiences of the
environment → without losing too much reward along the way.

• Exploration: finds more information about the environment.


• Exploitation: exploits known information to maximize reward.

• Interesting trade-off:
• immediate reward (exploitation) vs. gaining knowledge that might enable
higher future reward (exploration)
Exploration and Exploitation: Examples
• Restaurant Selection
• Exploitation: Go to your favorite restaurant
• Exploration: Try a new restaurant

• Oil Drilling
• Exploitation: Drill at the best-known location
• Exploration: Drill at a new location
Major Challenges of RL
• Sample efficiency
• RL algorithms require a large amount of data and experience to learn
effectively → costly and time-consuming.

• State and action spaces


• Exponential growth of state and action spaces as problem complexity
increases.
• For example, in the game of Go, the number of possible board configurations
is estimated to be 10170 .
Major Challenges of RL
• Exploration and exploitation
• Both exploration and exploitation are essential for learning, but they can also
conflict with each other.
Multi-Agent Reinforcement Learning
• Vanilla reinforcement learning is concerned with a single agent.

• Multi-agent reinforcement learning (MARL) studies how multiple


agents interact in a common environment.

• Cooperative: All agents working towards a common goal

• Competitive: Agents competing with one another to accomplish a goal


Multi-Agent Reinforcement Learning
Why MARL?
• A more natural decomposition of the problem: Need to train
policies for cellular antenna tilt control.
• Instead of training a single super-agent that controls all the cellular
antennas in a city→ it is more natural to model each antenna as a
separate agent in the environment.

• Potential for more scalable learning

Sing super-agent Multi-agent


MARL: Use case
• Dota 2: AI agents are trained to coordinate with each other to
compete against humans.
MARL: Challenges
• Environment non-stationarity

Fig: Non-stationarity of environment: Initially (a), the red agent learns to regulate
the speed of the traffic by slowing down. However, over time the blue agents learn
to bypass the red agent (b), rendering the previous experiences of the red agent
invalid.
References
• Reinforcement Learning: An Introduction by Richard S. Sutton and
Andrew G. Barto.
• David Silver Course on RL

You might also like