Reinforcement Learning
A simple way to understand 3rd paradigm of machine learning
Put simply, Reinforcement Learning (RL) is a type of machine learning that makes the agent discover or learn in an environment by using the trial and error method from its own actions by trying them. In most intriguing and challenging cases, actions might affect later.
Trial and error search and delayed reward are the two most important features of reinforcement learning.
“The learner is not told which actions to take, but instead must discover which actions yield the most reward by trying them.”
Topics whose name end with “ing” are simultaneously a problem and the field that studies the problem and its methods of solution. Such as Machine learning, Reinforcement learning Deep and Mountaineering, etc.
To understand reinforcement learning it is very important to know the difference between the problem and solution methods. A lot of people fail to understand the distinction and as result, they get confused and puzzled.
So, our basic task is to catch the important element of the real problem facing a learning subject interacting over a time period with given conditions to reach a goal.
A learning agent must be able to detect the state of its environment and must be able to take certain actions that might affect the state.
Now let us understand reinforcement learning in another way.
Reinforcement learning is different from Supervised Learning and Unsupervised Learning. These terms would seem to simplify machine learning paradigms. But they do not.
Supervised learning is learning from a set of labeled examples provided by some external observer. Each example or scenario is a description of a situation together with a pattern — The label — If a subject does some action the system should take directly to the situation where it belongs.
Unsupervised learning is learning which is about finding arrangements or structures hidden in collections of unlabeled data.
Now you might be thinking of reinforcement learning as some kind of Unsupervised learning because it does not depend on some labeled behavior.
Reinforcement learning is striving to extend the reward signal instead of trying to find hidden clues or structures.
Therefore, reinforcement learning is the 3rd paradigm of machine learning along with supervised, unsupervised, and other paradigms.
One of the most interesting aspects of modern reinforcement learning is its substantial interaction with main engineering, artificial intelligence, and other scientific disciplines.
Reinforcement learning is part of a long trend within machine learning and AI to greater association with statistics and other mathematical subjects.
Peculiarly, reinforcement learning has also strong interaction with neuroscience and psychology.
Of all the other types of machine learning, reinforcement learning is the closest one that humans and animals do and learning methods were inspired by biological learning systems such as trial and error mechanisms in humans and animals.
Examples and possible applications to understand reinforcement learning
A portable robot decides whether it should enter a new room to collect the trash or start going back to find its battery recharging point. The machine makes decisions based on the charge level of its battery and how easily and rapidly it has been able to locate the recharger.
A gazelle (Antelope) struggles to its feet minutes after being born. Half an hour later, it is running at 25 miles per hour.
A master chess player makes a move. The machine analyzes the opponent’s anticipating possible replies and counterreplies — do a judgment of the appropriateness of a particular position and move.
The above example share features that are so simple and easy to overlook. All involve interaction between the agent and environment, in which the agent looks to achieve the target without knowing the environment.
Agent’s action might affect the future state of the environment like the next chess position and the robot's next location.
Elements of reinforcement learning
After the agent and environment, you must be able to identify four main elements of the reinforcement learning system.
1. policy
2. reward signal
3. value function
4. model of the environment
The policy defines the learning agent’s way of behaving at a particular time. It relates to what in psychology would be called a set of stimuli — reaction rules or associations. The policy can be a simple task, or it may also involve extensive data such as a search process.
Reward Signal describes the goal of a reinforcement learning problem. The environment sends a signal number called reward. The agent's purpose is to maximize the reward it gets over the longer run. The reward signal, therefore, describes what are the good and bad actions of the agent.
The reward signal indicates what is good in an instant sensation, but a value function indicates what is better in the long run. In plain English, the value of a state is the total sum of rewards an agent can expect to collect over the future, starting from that state.
If we make a human analogy to simplify it even more, let’s say rewards are like pleasure (if high) and pain is (if low), although values relate to more superior or perceptive judgment of How happy or sad you are that your environment is in a particular state.
In fact, the most essential part of the whole reinforcement learning algorithms we think about is a method of perfectly estimating values.
Last but not least the final element of reinforcement learning systems is a model of the environment. It mirrors the behavior of the environment or allows judgment to be made about how the environment will behave.
Is reinforcement learning the future of machine learning?
Reinforcement learning, machine learning, and deep learning are interlinked with each other so no one of them is going to replace or dominates the others.
“There is a famous joke that reinforcement learning is the cherry on a great AI cake with machine learning cake itself and deep learning icing. Without the previous iterations, the cherry would top nothing,” Yann Lechun, the French scientist said.
Final Thought
Reinforcement learning is a cutting-edge modern technology that has the potential to change our world. It seems to be the most efficient way to make a machine innovative — seeking creative ways to do its tasks in a quick ingenuity.
However, reinforcement learning has the potential to be revolutionary technology and the next advancement in the world of AI
Rabindar Kumar
References:
Reinforcement Learning: An Introduction second edition Richard S. Sutton and Andrew G. Barto(A Summarize article from a book)
If you like this piece, share it with your buddies.