Week 1

Last updated Aug 10, 2022

Metadata

• URL:: Prediction and Control with Function Approximation | Coursera
• Tags:: #on/Reinforcement-Learning
• Links:: Reinforcement Learning Specialization
• Status:: #in-progress
• Instructor:: Martha White, Adam White
• Created:: 2022-06-06 18:23:17
• Modified:: 2022-06-06 18:23:10

About this Course
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 environment

• Tabular methods aren’t very applicable in real-world scenarios as they require to store data for each value.
• Function approximation methods, both linear and non-linear, are a good alternative to learn/estimate value functions.
  • Instead of storing the value of each state in a table, we only store a small set of weights that are used to estimate the value function.
• In case of non-linear methods, neural networks are used to approximate and learn the weights required to estimate the value function.
• Generalization intuitively means applying knowledge about specific situations to draw conclusions about a wider variety of situations.
  • In the context of policy evaluation, it means that updates to the value estimate of one state influence the value of other states.
  • It can speed-up learning by making better use of the experiences that we have.
• Discrimination means the ability to make the values for two states different to distinguish between the values for these two states.
• Tabular representations have provided good discrimination, but no generalization.
• It is ideal to have both high generalization and high discrimination, but in practice there’s typically a trade-off between the two.
• We can frame the policy evaluation task as a supervised learning problem.
  • But not all methods from supervised learning are ideal for reinforcement learning.
    • We need methods that are compatible with online-learning and bootstrapping.

• Policy evaluation under function approximation requires us to specify an objective.
• One possible objective is the Mean Squared Value Error $$\bar{VE} = \Sigma_s \mu(s) \left[ v_{\pi}(s) - \hat{v}(s, w)\right]^2.$$
• Gradient Descent and Stochastic Gradient Descent can be used to find stationary points of objectives.
  • These solutions are not always globally optimal.