- Level Professional
- Duration 22 hours
- Course by University of Alberta
-
Offered by
About
In this course, you will learn how to solve problems with large, high-dimensional, and potentially infinite state spaces. You will see that estimating value functions can be cast as a supervised learning problem---function approximation---allowing you to build agents that carefully balance generalization and discrimination in order to maximize reward. We will begin this journey by investigating how our policy evaluation or prediction methods like Monte Carlo and TD can be extended to the function approximation setting. You will learn about feature construction techniques for RL, and representation learning via neural networks and backprop. We conclude this course with a deep-dive into policy gradient methods; a way to learn policies directly without learning a value function. In this course you will solve two continuous-state control tasks and investigate the benefits of policy gradient methods in a continuous-action environment. Prerequisites: This course strongly builds on the fundamentals of Courses 1 and 2, and learners should have completed these before starting this course. Learners should also be comfortable with probabilities & expectations, basic linear algebra, basic calculus, Python 3.0 (at least 1 year), and implementing algorithms from pseudocode. By the end of this course, you will be able to: -Understand how to use supervised learning approaches to approximate value functions -Understand objectives for prediction (value estimation) under function approximation -Implement TD with function approximation (state aggregation), on an environment with an infinite state space (continuous state space) -Understand fixed basis and neural network approaches to feature construction -Implement TD with neural network function approximation in a continuous state environment -Understand new difficulties in exploration when moving to function approximation -Contrast discounted problem formulations for control versus an average reward problem formulation -Implement expected Sarsa and Q-learning with function approximation on a continuous state control task -Understand objectives for directly estimating policies (policy gradient objectives) -Implement a policy gradient method (called Actor-Critic) on a discrete state environmentModules
Course Introduction
1
Discussions
- Meet and Greet
2
Videos
- Course 3 Introduction
- Meet your instructors!
2
Readings
- Read Me: Pre-requisites and Learning Objectives
- Reinforcement Learning Textbook
Estimating values functions with supervised learning
3
Videos
- Moving to Parameterized Functions
- Generalization and Discrimination
- Framing Value Estimation as Supervised Learning
2
Readings
- Module 1 Learning Objectives
- Weekly Reading: On-policy Prediction with Approximation
The Objective for On-policy Prediction
4
Videos
- The Value Error Objective
- Introducing Gradient Descent
- Gradient Monte for Policy Evaluation
- State Aggregation with Monte Carlo
The Objective for TD
3
Videos
- Semi-Gradient TD for Policy Evaluation
- Comparing TD and Monte Carlo with State Aggregation
- Doina Precup: Building Knowledge for AI Agents with Reinforcement Learning
Linear TD
- Semi-gradient TD(0) with State Aggregation
1
Assignment
- On-policy Prediction with Approximation
1
Discussions
- Good Objectives for Control
3
Videos
- The Linear TD Update
- The True Objective for TD
- Week 1 Summary
Feature Construction for Linear Methods
4
Videos
- Coarse Coding
- Generalization Properties of Coarse Coding
- Tile Coding
- Using Tile Coding in TD
2
Readings
- Module 2 Learning Objectives
- Weekly Reading: On-policy Prediction with Approximation II
Neural Networks
3
Videos
- What is a Neural Network?
- Non-linear Approximation with Neural Networks
- Deep Neural Networks
Training Neural Networks
- Semi-gradient TD with a Neural Network
1
Assignment
- Constructing Features for Prediction
1
Discussions
- Constructing Features for Prediction
4
Videos
- Gradient Descent for Training Neural Networks
- Optimization Strategies for NNs
- David Silver on Deep Learning + RL = AI?
- Week 2 Review
Episodic Sarsa with Function Approximation
3
Videos
- Episodic Sarsa with Function Approximation
- Episodic Sarsa in Mountain Car
- Expected Sarsa with Function Approximation
2
Readings
- Module 3 Learning Objectives
- Weekly Reading: On-policy Control with Approximation
Exploration under Function Approximation
1
Videos
- Exploration under Function Approximation
Average Reward
- Function Approximation and Control
1
Assignment
- Control with Approximation
2
Discussions
- Control with FA #1
- Control with FA #2
3
Videos
- Average Reward: A New Way of Formulating Control Problems
- Satinder Singh on Intrinsic Rewards
- Week 3 Review
Learning Parameterized Policies
2
Videos
- Learning Policies Directly
- Advantages of Policy Parameterization
2
Readings
- Module 4 Learning Objectives
- Weekly Reading: Policy Gradient Methods
Policy Gradient for Continuing Tasks
2
Videos
- The Objective for Learning Policies
- The Policy Gradient Theorem
Actor-Critic for Continuing Tasks
2
Videos
- Estimating the Policy Gradient
- Actor-Critic Algorithm
Policy Parameterizations
- Average Reward Softmax Actor-Critic using Tile-coding
1
Assignment
- Policy Gradient Methods
1
Discussions
- Policy Gradient methods
4
Videos
- Actor-Critic with Softmax Policies
- Demonstration with Actor-Critic
- Gaussian Policies for Continuous Actions
- Week 4 Summary
Course Wrap-up
1
Videos
- Congratulations! Course 4 Preview
Auto Summary
Unlock the power of advanced problem-solving in the realm of large, high-dimensional, and infinite state spaces with the "Prediction and Control with Function Approximation" course, a part of Coursera's Data Science & AI offerings. Led by expert instructors, this professional-level course teaches you how to leverage supervised learning to approximate value functions, equipping you to build intelligent agents that effectively balance generalization and discrimination to maximize rewards. Throughout the duration of the course, you'll explore the extension of policy evaluation methods like Monte Carlo and Temporal Difference (TD) to function approximation settings. You'll dive deep into feature construction techniques for reinforcement learning (RL) and representation learning via neural networks and backpropagation. The course culminates in a thorough examination of policy gradient methods, enabling you to learn policies directly without needing a value function. Key learning outcomes include: - Mastering supervised learning for value function approximation - Implementing TD with function approximation in continuous state environments - Constructing features using fixed basis and neural networks - Tackling the challenges of exploration in function approximation - Distinguishing between discounted and average reward problem formulations for control - Implementing advanced methods like expected Sarsa, Q-learning, and policy gradient methods Ideal for those with a solid foundation in probabilities, linear algebra, calculus, and Python programming (at least one year), this course builds on previous coursework and requires comfort with implementing algorithms from pseudocode. This comprehensive 1320-minute course is available through Coursera's Starter subscription, making it accessible for learners looking to elevate their expertise in data science and AI. Join now to gain cutting-edge skills in prediction and control with function approximation, and take your professional capabilities to new heights.

Martha White

Adam White