2022-11-13

Reinforcement Learning

Table of Contents

a_taxonomy_of_rl_algorithms-20230727194644.png

Figure 1: A Taxonomy of RL Algorithms [from spinningup.openai.com]

1. Resources

Lectures:

Books:

  • Reinforcement Learning: An Introduction by Sutton and Barto is still the intro book that estabilishes base for learning RL.
  • Distributional Reinforcement Learning by Bellemare, Dabney & Rowland is a specialied book looking at RL in a new perspective that advances the theory of RL
  • Deep Reinforcement Learning Hands-On by Maxim Lapan (2024) is practical and implementation focused approach
  • Multi-Agent Reinforcement Learning: Foundations and Modern Approaches by Albrecht, Christianos & Schafer (2024) gives comprehensive intro to MARL

Software:

Articles:

Podcasts:

2. Multi Agent Reinforcement Learning

3. Papers

4. Concepts

4.2. True Gradient Descent

  • Is DQN True Gradient Descent? No. They are approximations to it. CS234 - Lecture 6 (t=3253)

    Because it is not a derivative of a loss function.

  • GTD (Gradient Temporal Difference) are true gradient descent
  • The first section of chapter on Function Approximation (Sutton & Barto) has few points on this.

4.3. DQN

4.3.1. Replay Buffer

4.3.2. Stable Target

4.3.3. Prioritized Replay Buffer

4.3.4. Double DQN

4.3.5. Duelling DQN

4.3.6. Maximization Bias

  • Double DQN
  • Stable Target

prevent maximization bias

4.4. Actions are Infinite

From Assumptions of Decision Making Models in AGI:

It is unreasonable to assume that at any state, all possible actions are listed.

  • Actions in small scale may be discrete or a finite collection of distributions, but at the level where planning happens set of all possible actions is infinite.

    Such actions can in principle to thought to be recursively composed of a set of basic operations/actions. But the decision making happens not at those basic actions but at level of composed actions.

  • It is a task in itself to know what actions can be taken and what actions should we evaluate.
  • Thus decision making involves composing short timestep actions to get longer term action over which planning can be done.

    i.e. decision making is often not about selection but selective composition. [Page 2]

So, one thing to explore would be Hierarchical Reinforcement Learning.

4.5. RL Without Rewards

What does this even mean?

Is it even RL if there are no rewards. So, this must be limited to improving performance in downstream tasks that do indeed have rewards?

What is the utility?

  • helps when the reward is sparse
  • skills developed without reward can act as primitive for hierarchical RL

Work done so far:

  • ICVF
  • Diversity is all you need.

Backlinks


You can send your feedback, queries here