Consider that you are teaching the dog to catch a ball, but you cannot teach the dog explicitly to catch a ball; instead, you will just throw a ball, and every time the dog catches the ball, you will give it a cookie. If it fails to catch the ball, you will not give a cookie. The dog will figure out what actions made it receive a cookie and will repeat those actions.

Similarly, in a RL environment, you will not teach the agent what to do or how to do instead, you will give a reward to the agent for each action it does. The reward may be positive or negative. Then the agent will start performing actions which made it receive a positive reward. Thus, it is a trial and error process. In the previous analogy, the dog represents the agent. Giving a cookie to the dog upon catching the ball is a positive reward, and not giving a cookie is a negative reward.

There might be delayed rewards. You may not get a reward at each step. A reward may be given only after the completion of a task. In some cases, you get a reward at each step to find out that whether you are making any mistakes.

Imagine you want to teach a robot to walk without getting stuck by hitting a mountain, but you will not explicitly teach the robot not to go in the direction of the mountain:

Instead, if the robot hits and get stuck on the mountain, you will take away ten points so that robot will understand that hitting the mountain will result in a negative reward and it will not go in that direction again:

You will give 20 points to the robot when it walks in the right direction without getting stuck. So the robot will understand which is the right path and will try to maximize the rewards by going in the right direction:

The RL agent can explore different actions which might provide a good reward or it can exploit (use) the previous action which resulted in a good reward. If the RL agent explores different actions, there is a great possibility that the agent will receive a poor reward as all actions are not going to be the best one. If the RL agent exploits only the known best action, there is also a great possibility of missing out on the best action, which might provide a better reward. There is always a trade-off between exploration and exploitation. We cannot perform both exploration and exploitation at the same time. We will discuss the exploration-exploitation dilemma in detail in the upcoming chapters.

The steps involved in typical RL algorithm are as follows:

1. First, the agent interacts with the environment by performing an action
2. The agent performs an action and moves from one state to another
3. And then the agent will receive a reward based on the action it performed
4. Based on the reward, the agent will understand whether the action was good or bad
5. If the action was good, that is, if the agent received a positive reward, then the agent will prefer performing that action or else the agent will try performing an other action which results in a positive reward. So it is basically a trial and error learning process

In supervised learning, the machine (agent) learns from training data which has a labeled set of input and output. The objective is that the model extrapolates and generalizes its learning so that it can be well applied to the unseen data. There is an external supervisor who has a complete knowledge base of the environment and supervises the agent to complete a task.

Consider the dog analogy we just discussed; in supervised learning, to teach the dog to catch a ball, we will teach it explicitly by specifying turn left, go right, move forward five steps, catch the ball, and so on. But instead in RL we just throw a ball, and every time the dog catches the ball, we give it a cookie (reward). So the dog will learn to catch the ball that meant it received a cookie.

In unsupervised learning, we provide the model with training data which only has a set of inputs; the model learns to determine the hidden pattern in the input. There is a common misunderstanding that RL is a kind of unsupervised learning but it is not. In unsupervised learning, the model learns the hidden structure whereas in RL the model learns by maximizing the rewards. Say we want to suggest new movies to the user. Unsupervised learning analyses the similar movies the person has viewed and suggests movies, whereas RL constantly receives feedback from the user, understands his movie preferences, and builds a knowledge base on top of it and suggests a new movie.

There is also another kind of learning called semi-supervised learning which is basically a combination of supervised and unsupervised learning. It involves function estimation on both the labeled and unlabeled data, whereas RL is essentially an interaction between the agent and its environment. Thus, RL is completely different from all other machine learning paradigms.

The elements of RL are shown in the following sections.

Agents are the software programs that make intelligent decisions and they are basically learners in RL. Agents take action by interacting with the environment and they receive rewards based on their actions, for example, Super Mario navigating in a video game.

A policy defines the agent's behavior in an environment. The way in which the agent decides which action to perform depends on the policy. Say you want to reach your office from home; there will be different routes to reach your office, and some routes are shortcuts, while some routes are long. These routes are called policies because they represent the way in which we choose to perform an action to reach our goal. A policy is often denoted by the symbol 𝛑. A policy can be in the form of a lookup table or a complex search process.

A value function denotes how good it is for an agent to be in a particular state. It is dependent on the policy and is often denoted by v(s). It is equal to the total expected reward received by the agent starting from the initial state. There can be several value functions; the optimal value function is the one that has the highest value for all the states compared to other value functions. Similarly, an optimal policy is the one that has the optimal value function.

Model is the agent's representation of an environment. The learning can be of two types—model-based learning and model-free learning. In model-based learning, the agent exploits previously learned information to accomplish a task, whereas in model-free learning, the agent simply relies on a trial-and-error experience for performing the right action. Say you want to reach your office from home faster. In model-based learning, you simply use a previously learned experience (map) to reach the office faster, whereas in model-free learning you will not use a previous experience and will try all different routes and choose the faster one.

Agents are the software agents that perform actions, A_t, at a time, t, to move from one state, S_t, to another state S_t+1. Based on actions, agents receive a numerical reward, R, from the environment. Ultimately, RL is all about finding the optimal actions that will increase the numerical reward:

Let us understand the concept of RL with a maze game:

The objective of a maze is to reach the destination without getting stuck on the obstacles. Here's the workflow:

• The agent is the one who travels through the maze, which is our software program/ RL algorithm
• The environment is the maze
• The state is the position in a maze that the agent currently resides in
• An agent performs an action by moving from one state to another
• An agent receives a positive reward when its action doesn't get stuck on any obstacle and receives a negative reward when its action gets stuck on obstacles so it cannot reach the destination
• The goal is to clear the maze and reach the destination

Everything agents interact with is called an environment. The environment is the outside world. It comprises everything outside the agent. There are different types of environment, which are described in the next sections.

An environment is said to be deterministic when we know the outcome based on the current state. For instance, in a chess game, we know the exact outcome of moving any player.

An environment is said to be stochastic when we cannot determine the outcome based on the current state. There will be a greater level of uncertainty. For example, we never know what number will show up when throwing a dice.

When an agent can determine the state of the system at all times, it is called fully observable. For example, in a chess game, the state of the system, that is, the position of all the players on the chess board, is available the whole time so the player can make an optimal decision.

When an agent cannot determine the state of the system at all times, it is called partially observable. For example, in a poker game, we have no idea about the cards the opponent has.

When there is only a finite state of actions available for moving from one state to another, it is called a discrete environment. For example, in a chess game, we have only a finite set of moves.

When there is an infinite state of actions available for moving from one state to another, it is called a continuous environment. For example, we have multiple routes available for traveling from the source to the destination.

The episodic environment is also called the non-sequential environment. In an episodic environment, an agent's current action will not affect a future action, whereas in a non-episodic environment, an agent's current action will affect a future action and is also called the sequential environment. That is, the agent performs the independent tasks in the episodic environment, whereas in the non-episodic environment all agents' actions are related.

As the names suggest, a single-agent environment has only a single agent and the multi-agent environment has multiple agents. Multi-agent environments are extensively used while performing complex tasks. There will be different agents acting in completely different environments. Agents in a different environment will communicate with each other. A multi-agent environment will be mostly stochastic as it has a greater level of uncertainty.

RL platforms are used for simulating, building, rendering, and experimenting with our RL algorithms in an environment. There are many different RL platforms available, as described in the next sections.

OpenAI Gym is a toolkit for building, evaluating, and comparing RL algorithms. It is compatible with algorithms written in any framework like TensorFlow, Theano, Keras, and so on. It is simple and easy to comprehend. It makes no assumption about the structure of our agent and provides an interface to all RL tasks.

OpenAI Universe is an extension to OpenAI Gym. It provides an ability to train and evaluate agents on a wide range of simple to real-time complex environments. It has unlimited access to many gaming environments. Using Universe, any program can be turned into a Gym environment without access to program internals, source code, or APIs as Universe works by launching the program automatically behind a virtual network computing remote desktop.

DeepMind Lab is another amazing platform for AI agent-based research. It provides a rich simulated environment that acts as a lab for running several RL algorithms. It is highly customizable and extendable. The visuals are very rich, science fiction-style, and realistic.

RL-Glue provides an interface for connecting agents, environments, and programs together even if they are written in different programming languages. It has the ability to share your agents and environments with others for building on top of your work. Because of this compatibility, reusability is greatly increased.

Project Malmo is the another AI experimentation platform from Microsoft which builds on top of Minecraft. It provides good flexibility for customizing the environment. It is integrated with a sophisticated environment. It also allows overclocking, which enables programmers to play out scenarios faster than in standard Minecraft. However, Malmo currently only provides Minecraft gaming environments, unlike Open AI Universe.

ViZDoom, as the name suggests, is a doom-based AI platform. It provides support for multi-agents and a competitive environment to test the agent. However, ViZDoom only supports the Doom game environment. It provides off-screen rendering and single and multiplayer support.

With greater advancements and research, RL has rapidly evolved everyday applications in several fields ranging from playing computer games to automating a car. Some of the RL applications are listed in the following sections.

Many online education platforms are using RL for providing personalized content for each and every student. Some students may learn better from video content, some may learn better by doing projects, and some may learn better from notes. RL is used to tune educational content personalized for each student according to their learning style and that can be changed dynamically according to the behavior of the user.

RL has endless applications in medicine and health care; some of them include personalized medical treatment, diagnosis based on a medical image, obtaining treatment strategies in clinical decision making, medical image segmentation, and so on.

In manufacturing, intelligent robots are used to place objects in the right position. If it fails or succeeds in placing the object at the right position, it remembers the object and trains itself to do this with greater accuracy. The use of intelligent agents will reduce labor costs and result in better performance.

RL is extensively used in inventory management, which is a crucial business activity. Some of these activities include supply chain management, demand forecasting, and handling several warehouse operations (such as placing products in warehouses for managing space efficiently). Google researchers in DeepMind have developed RL algorithms for efficiently reducing the energy consumption in their own data center.

RL is widely used in financial portfolio management, which is the process of constant redistribution of a fund into different financial products and also in predicting and trading in commercial transactions markets. JP Morgan has successfully used RL to provide better trade execution results for large orders.

With the unified power of deep learning and RL, Deep Reinforcement Learning (DRL) has been greatly evolving in the fields of Natural Language Processing (NLP) and Computer Vision (CV). DRL has been used for text summarization, information extraction, machine translation, and image recognition, providing greater accuracy than current systems.

In this chapter, we have learned the basics of RL and also some key concepts. We learned different elements of RL and different types of RL environments. We also covered the various available RL platforms and also the applications of RL in various domains.

In the next chapter, Chapter 2, Getting Started with OpenAI and TensorFlow, we will learn the basics of and how to install OpenAI and TensorFlow, followed by simulating environments and teaching the agents to learn in the environment.

The question list is as follows:

1. What is reinforcement learning?
2. How does RL differ from other ML paradigms?
3. What are agents and how do agents learn?
4. What is the difference between a policy function and a value function?
5. What is the difference between model-based and model-free learning?
6. What are all the different types of environments in RL?
7. How does OpenAI Universe differ from other RL platforms?
8. What are some of the applications of RL?

Overview of RL: https://www.cs.ubc.ca/~murphyk/Bayes/pomdp.html.