Original article was published by Thomas Simonini on Artificial Intelligence on Medium
This article is part of Deep Reinforcement Learning Course. A free course from beginner to expert. Check the syllabus here.
Welcome to the most fascinating topic in Artificial Intelligence: Deep Reinforcement Learning.
Deep RL is a type of Machine Learning where an agent learns how to behave in an environment by performing actions and seeing the results.
Since 2013 and the Deep Q-Learning paper, we’ve seen a lot of breakthroughs. From OpenAI five that beat some of the best Dota2 players of the world, to the Dexterity project, we live in an exciting moment in Deep RL research.
Moreover, since the first version of this course in 2018, a ton of new libraries (TF-Agents, Stable-Baseline 2.0…) and environments where launched: MineRL (Minecraft), Unity ML-Agents, OpenAI retro (NES, SNES, Genesis games…). You have now access to so many amazing games to build your agents.
That’s why this is the best moment to start learning, and with this course you’re in the right place.
Yes, because this article is the first chapter of Deep RL Course v2.0 a free course from beginner to expert where you’ll master the skills and architectures you need, to become a deep reinforcement learning expert.
During this course, you’ll build a strong professional portfolio by implementing awesome agents with Tensorflow and PyTorch that learn to play Space invaders, Minecraft, Starcraft, Sonic the hedgehog and more!
So in this first chapter, you’ll learn the foundations of deep reinforcement learning.
It’s really important to master these elements before diving into implementing Deep Reinforcement Learning agents. The goal in this chapter is to give you solid foundations.
If you prefer, you can watch the 📹 video version of this chapter:
So let’s get started!
What is Reinforcement Learning?
In order to understand what is reinforcement learning, let’s start with the big picture.
The big picture
The idea behind Reinforcement Learning is that an agent (an AI) will learn from the environment by interacting with it (through trial and error) and receiving rewards (negative or positive) as feedback for performing actions.
Learning from interaction with the environment comes from our natural experiences.
For instance, imagine you put your little brother in front of a video game he never played, a controller in his hands, and let him alone.
Your brother will interact with the environment (the video game) by pressing the right button (action). He got a coin, that’s a +1 reward. It’s positive, he just understood that in this game he must get the coins.
But then, he presses right again and he touches an enemy, he just died -1 reward.
By interacting with his environment through trial and error, your little brother just understood that in this environment, he needs to get coins, but avoid the enemies.
Without any supervision, the child will get better and better at playing the game.
That’s how humans and animals learn, through interaction. Reinforcement Learning is just a computational approach of learning from action.
A formal definition
If we take now a formal definition:
Reinforcement learning is a framework for solving control tasks (also called decision problems) by building agents that learn from the environment by interacting with it through trial and error and receiving rewards (positive or negative) as unique feedback.
But how Reinforcement Learning works?
The Reinforcement Learning Framework
The RL Process
To understand the RL process, let’s imagine an agent learning to play a platform game:
- Our Agent receives state S0 from the Environment — we receive the first frame of our game (environment).
- Based on that state S0, the agent takes an action A0 — our agent will move to the right.
- Environment transitions to a new state S1 — new frame.
- Environment gives some reward R1 to the agent — we’re not dead (Positive Reward +1).
This RL loop outputs a sequence of state, action and reward and next state.
The goal of the agent is to maximize its cumulative reward, called the expected return.
The reward hypothesis: the central idea of Reinforcement Learning
Why the goal of the agent is to maximize the expected return?
Because RL is based on the reward hypothesis, which is that all goals can be described as the maximization of the expected return (expected cumulative reward).
That’s why in Reinforcement Learning, to have the best behavior, we need to maximize the expected cumulative reward.
(Optional) Markov Property
You’ll see in papers that the RL process is called the Markov Decision Process (MDP).
We’ll talk again about the Markov Property in the next chapters. But if you need to remember something today about it is just that Markov Property implies that our agent needs only the current state to make its decision about what action to take and not the history of all the states and actions he took before.
Now let’s dive a little bit on all this new vocabulary:
Observations/States are the information our agent gets from the environment. In the case of a video game, it can be a frame (a screenshot), in the case of the trading agent, it can be the value of a certain stock etc.
There is a differentiation to make between observation and state:
- State s: is a complete description of the state of the world (there is no hidden information). In a fully observed environment.
With a chess game, we are in a fully observed environment, since we have access to the whole check board information.
- Observation o: is a partial description of the state. In a partially observed environment.
In Super Mario Bros, we are in a partially observed environment, we receive an observation since we only see a part of the level.
In reality, we use the term state in this course but we will make the distinction in implementations.
The Action space is the set of all possible actions in an environment.
The actions can come from a discrete or continuous space:
- Discrete space: the number of possible actions is finite.
In Super Mario Bros, we have a finite set of actions since we have only 4 directions and jump.
- Continuous space: the number of possible actions is infinite.
A Self Driving Car agent has an infinite number of possible actions since he can turn left 20°, 21°, 22°, honk, turn right 20°, 20,1°…
Taking this information into consideration is crucial because it will have importance when we will choose in the future the RL algorithm.
Rewards and the discounting
The reward is fundamental in RL because it’s the only feedback for the agent. Thanks to it, our agent knows if the action taken was good or not.
The cumulative reward at each time step t can be written as:
Which is equivalent to:
However, in reality, we can’t just add them like that. The rewards that come sooner (at the beginning of the game) are more probable to happen, since they are more predictable than the long term future reward.
Let say your agent is this small mouse that can move one tile each time step, and your opponent is the cat (that can move too). Your goal is to eat the maximum amount of cheese before being eaten by the cat.
As we can see in the diagram, it’s more probable to eat the cheese near us than the cheese close to the cat (the closer we are to the cat, the more dangerous it is).
As a consequence, the reward near the cat, even if it is bigger (more cheese), will be more discounted since we’re not really sure we’ll be able to eat it.
To discount the rewards, we proceed like this:
- We define a discount rate called gamma. It must be between 0 and 1.
- The larger the gamma, the smaller the discount. This means our agent cares more about the long term reward.
- On the other hand, the smaller the gamma, the bigger the discount. This means our agent cares more about the short term reward (the nearest cheese).
2. Then, each reward will be discounted by gamma to the exponent of the time step.
As the time step increases, the cat gets closer to us, so the future reward is less and less probable to happen.
Our discounted cumulative expected rewards is:
Type of tasks
A task is an instance of a Reinforcement Learning problem. We can have two types of tasks: episodic and continuous.
In this case, we have a starting point and an ending point (a terminal state). This creates an episode: a list of States, Actions, Rewards, and New States.
For instance think about Super Mario Bros, an episode begin at the launch of a new Mario Level and ending when you’re killed or you’re reach the end of the level.
These are tasks that continue forever (no terminal state). In this case, the agent has to learn how to choose the best actions and simultaneously interacts with the environment.
For instance, an agent that do automated stock trading. For this task, there is no starting point and terminal state. The agent keeps running until we decide to stop him.
Exploration/ Exploitation tradeoff
Finally, before looking at the different methods to solve Reinforcement Learning problems, we must cover one more very important topic: the exploration/exploitation trade-off.
- Exploration is exploring the environment by trying random actions in order to find more information about the environment.
- Exploitation is exploiting known information to maximize the reward.
Remember, the goal of our RL agent is to maximize the expected cumulative reward. However, we can fall into a common trap.
Let’s take an example:
In this game, our mouse can have an infinite amount of small cheese (+1 each). But at the top of the maze, there is a gigantic sum of cheese (+1000).
However, if we only focus on exploitation, our agent will never reach the gigantic sum of cheese. Instead, it will only exploit the nearest source of rewards, even if this source is small (exploitation).
But if our agent does a little bit of exploration, it can discover the big reward (the pile of big cheese).
This is what we call the exploration/exploitation trade off. We need to balance how much we explore the environment and how much we exploit what we know about the environment.
Therefore, we must define a rule that helps to handle this trade-off. We’ll see in future chapters different ways to handle it.
If it’s still confusing think of a real problem: the choice of a restaurant:
Source: Berkley AI Course
- Exploitation: You go everyday to the same one that you know is good and take the risk to miss another better restaurant.
- Exploration: Try restaurants you never went before, with the risk of having a bad experience but the probable opportunity of an amazing experience.
The two main approaches for solving RL problems
Now that we learned the RL framework, how do we solve the RL problem?
In other terms, how to build a RL agent that can select the actions that maximize its expected cumulative reward?
The Policy π: the agent’s brain
The Policy π is the brain of our Agent, it’s the function that tell us what action to take given the state we are. So it defines the agent behavior at a given time.
This Policy is the function we want to learn, our goal is to find the optimal policy π*, the policy that maximizes expected return when the agent acts according to it. We find this π* through training.
There are two approaches to train our agent to find this optimal policy π*:
- Directly, by teaching the agent to learn which action to take, given the state is in: Policy-Based Methods.
- Indirectly, teach the agent to learn which state is more valuable and then take the action that leads to the more valuable states: Value-Based Methods.
In Policy-Based Methods, we learn a policy function directly.
This function will map from each state to the best corresponding action at that state. Or a probability distribution over the set of possible actions at that state.
We have two types of policy:
- Deterministic: a policy at a given state will always return the same action.
- Stochastic: output a probability distribution over actions.
Value based methods
In Value based methods, instead of training a policy function, we train a value function that maps a state to the expected value of being at that state.
The value of a state is the expected discounted return the agent can get if it starts in that state, and then act according to our policy.
“Act according to our policy” just means that our policy is “going to the state with the highest value”.
Here we see that our value function defined value for each possible state.
The “Deep” in Reinforcement Learning
Wait… you spoke about Reinforcement Learning, but why we speak about Deep Reinforcement Learning?
Deep Reinforcement Learning introduces deep neural networks to solve Reinforcement Learning problems — hence the name “deep.”
For instance, in the next article, we’ll work on Q-Learning (classic Reinforcement Learning) and then Deep Q-Learning both are value-based RL algorithms.
You’ll see the difference is that in the first approach, we use a traditional algorithm to create a Q table that helps us find what action to take for each state.
In the second approach, we will use a Neural Network (to approximate the q value).
If you are not familiar with Deep Learning you definitely should watch the MIT Intro Course on Deep Learning (Free)
That was a lot of information, if we summarize:
- Reinforcement Learning is a computational approach of learning from action. We build an agent that learns from the environment by interacting with it through trial and error and receiving rewards (negative or positive) as feedback.
- The goal of any RL agent is to maximize its expected cumulative reward (also called expected return) because RL is based on the reward hypothesis, which is that all goals can be described as the maximization of the expected cumulative reward.
- The RL process is a loop that outputs a sequence of state, action, reward and next state.
- To calculate the expected cumulative reward (expected return), we discount the rewards: the rewards that come sooner (at the beginning of the game) are more probable to happen since they are more predictable than the long term future reward.
- To solve an RL problem, you want to find an optimal policy, the policy is the “brain” of your AI that will tell us what action to take given a state. The optimal one is the one who gives you the actions that max the expected return.
- There are two ways to find your optimal policy:
- By training your policy directly: policy-based methods.
- By training a value function that tells us the expected return the agent will get at each state and use this function to define our policy: value-based methods.
- Finally, we speak about Deep RL because we introduces deep neural networks to estimate the action to take (policy based) or to estimate the value of a state (value based) hence the name “deep.”