Reinforcement Learning (English): Master the Art of RL

Howdy and welcome to our course; Reinforcement Studying.

Reinforcement Studying is a really thrilling and essential subject of Machine Studying and AI. Some name it the crown jewel of AI.

On this course, we are going to cowl all of the features associated to Reinforcement Studying or RL. We’ll begin by defining the RL drawback, and evaluate it to the Supervised Studying drawback, and uncover the areas of functions the place RL can excel. This contains the issue formulation, ranging from the very fundamentals to the superior utilization of Deep Studying, resulting in the period of Deep Reinforcement Studying.

In our journey, we are going to cowl, as common, each the theoretical and sensible features, the place we are going to discover ways to implement the RL algorithms and apply them to the well-known issues utilizing libraries like OpenAI Gymnasium, Keras-RL, TensorFlow Brokers or TF-Brokers and Secure Baselines.

The course is split into 6 predominant sections:

1- We begin with an introduction to the RL drawback definition, primarily evaluating it to the Supervised studying drawback, and discovering the appliance domains and the primary constituents of an RL drawback. We describe right here the well-known OpenAI Gymnasium environments, which might be our playground in relation to sensible implementation of the algorithms that we study.

2- Within the second half we talk about the primary formulation of an RL drawback as a Markov Choice Course of or MDP, with easy answer to essentially the most fundamental issues utilizing Dynamic Programming.

3- After being armed with an understanding of MDP, we transfer on to discover the answer area of the MDP drawback, and what the completely different options past DP, which incorporates model-based and model-free options. We’ll focus on this half on model-free options, and defer model-based options to the final half. On this half, we describe the Monte-Carlo and Temporal-Distinction sampling based mostly strategies, together with the well-known and essential Q-learning algorithm, and SARSA. We’ll describe the sensible utilization and implementation of Q-learning and SARSA on management tabular maze issues from OpenAI Gymnasium environments.

4- To maneuver past easy tabular issues, we might want to study perform approximation in RL, which ends up in the mainstream RL strategies at present utilizing Deep Studying, or Deep Reinforcement Studying (DRL). We’ll describe right here the breakthrough algorithm of DeepMind that solved the Atari video games and AlphaGO, which is Deep Q-Networks or DQN. We additionally talk about how we are able to resolve Atari video games issues utilizing DQN in observe utilizing Keras-RL and TF-Brokers.

5- Within the fifth half, we transfer to Superior DRL algorithms, primarily beneath a household referred to as Coverage based mostly strategies. We talk about right here Coverage Gradients, DDPG, Actor-Critic, A2C, A3C, TRPO and PPO strategies. We additionally talk about the essential Secure Baseline library to implement all these algorithms on completely different environments in OpenAI Gymnasium, like Atari and others.

6- Lastly, we discover the model-based household of RL strategies, and importantly, differentiating model-based RL from planning, and exploring the entire spectrum of RL strategies.

Hopefully, you get pleasure from this course, and discover it helpful.