Introduction

We may have heard of the Machine Learning before, which is a subdiscipline of the technology of Artificial Intelligence (AI), which uses algorithm to make machines to acquire new sets of data and/or information over time through experiences and relationships with other systems. For a baseline knowledge of Machine Learning, you can read this post, so I wouldn't bore us with vain repetitions. In this survey, we will be looking at a step further from the Machine Learning, and this is called the Reinforcement Learning (RL).

_{[Image of Reinforcement Learning Interacting with the environment. Source: Wikimedia Commons. Author: Notfruit. CC BY-SA 4.0 Licensed]}

As a way of set-induction, I would give a brief definition of Reinforcement Learning; but for a clearer understanding; I would break the term apart: "Reinforcement" and "Learning". Reinforcement is simply the augmentation or strengthening of a particular thing, while "learning" is the process of acquiring new information. So putting the both of them in apposition with each other, you would simply say that; "Reinforcement Learning" (RL) is a type of Machine Learning that causes computer programs; or machines in general; to take decisions or perform an action in response to some changes in the environment, in order to yield a particular result.

Hard to understand? Okay let me use this scenario to explain further: In the Game of Go; which is a board game; each of the two human players would employ different combination strategies just to gain more territories. With the introduction of the AI Google's DeepMind AlphaGo, it was seen to beat the human Go champion. But how did this happen? The AlphaGo was using complex algorithms to detect lists of possible moves to counter the moves by the human Go champion, and employing the same to beat him.

By inference; Reinforcement Learning is simply a Machine Learning approach that causes the machine to acquire knowledge through its own actions and also through interactions. For example: As a kid; your parents may probably not teach you that fire burns. But the moment you stick your hand in fire and get some little nasty burns, you would quickly withdraw your hand. But not just that, you would associate "fire" with harm, and store up the info in your brain for future reference, and thus, avoiding contact with fire again. In this way, you have acquired the knowledge by your own actions. This is exactly what the Reinforcement Learning seeks to simulate.

_{[Image Homemade by me @samminator]}

Before the era of Machine Learning; almost every set of instructions would be explicitly coded into the machines by the programmers. But with the advent of Machine Learning, machines can now figure out ways to get things done of which they had not been pre-programmed to. This learning process could be in form of mimicking the natural process of behavioral psychology; just like the example I gave about "fire and hand".

It might interest you to know that; even though the technology of Reinforcement Learning is grouped as emerging technology (at least; for the complex ones), it had existed years back in the crude way. According to this report, in 1951, a particular machine was designed, which is the SNARC (Stochastic Neural Analogy Reinforcement Calculator) by Marvin Minsky who was considered one of the fathers of AI. This machine simulated the action of a rat and was able to learn how to navigate through a complex maze. At that time, it wasn't recognized as Reinforcement Learning, but on a closer observation, you would agree with me that; for the SNARC to navigate effectively through the maze; it would relate with tons of possible navigation routes. And each time it doesn't make a head-way, it would eliminate that particular route as a possible exit route, until it finally finds its way out of the maze. If you don't call this Reinforcement Learning, I wonder what you would call it.

Actors in Reinforcement Learning

There are some technologies that the Reinforcement Learning would hinge on to achieve the maximum effective result. One of these technologies is the Deep Learning (DL). Deep Learning is also a subdiscipline of Machine Learning; and is also closely related to the RL. But in the Deep Learning model, the machine seeks to be more concerned with recognizing patterns for similarly larger subsets of data, but RL relates with specialized data set (which can also be derived from the DL). So you see; the DL and the RL work hand-in-hand.

_{[Image Source: Wikimedia Commons. Author: Megajuice. Public Domain Licensed]}

Let's take this scenario of the autonomous driving vehicle to buttress some points: The purpose of the autonomous vehicle is to simulate the human driver right? But what about a case where the car is driving along a busy street and a pedestrian runs off on the road? First of all; it should relate to the fact that it shouldn't knock down the pedestrian. But to complicate the matters; the right side of the road is filled with people, and houses are on the left side. Looking at this; continuing straight would knock down the person, swerving to the right would be more catastrophic, and to the left would cause an impact with the building.

In the above scenario, the car would need an immediate solution to the complex problem by correlating with some possibilities of instant actions to circumvent the impending catastrophe. And remember, these actions may not be what have been pre-programmed into it; and it would require split seconds to effect the action. But prior to this; if the autonomous vehicle had passed through a similar computer simulations like; "racing with other drivers in a very busy and chaotic environment", it would be able to relate the experiences it had acquired, thus making the required decision in real-time.

A particular company; Mobileye has designed a software to achieve this, and it is called "Highway Merging". The purpose of this is to bring Reinforcement Learning to autonomous vehicles. This software has already been passed through some real-life simulations, and the data from these are merged into a single whole, so that the autonomous vehicles would relate with them to make quick decision as they present themselves. So in the scenario I gave above, the car could decide to jostle between the building and even brush against some part of the building just to save the pedestrian, and also to avoid running into other people on the right side.

Downturns in RL

There are some limitations in the technology of RL which scientists are figuring out ways to overcome. Most importantly, for an efficient learning process, there is a need for an extremely large data set (just like is applicable in the Deep Learning), and these data need to be gotten from the simulations of real-life scenarios. And dealing with such a large data would require machines of huge computational capacities. But thanks to emerging technologies; a solution has been provided, and the Digital Twinning technology can be incorporated into the RL.

_{[Image of Neural Network. Source: Wikimedia Commons. Author: Chrislb. CC BY-SA 3.0 Licensed]}

Also, with the technology of Neural Networks, complex instructions can be processed from extremely large sets of data. This is very needful, so that; for example; the autonomous driving car would not create an accident in the bid to avoid an impending accident.

Conclusion

We have seen another step ahead in the field of Machine Learning, and that is the Reinforcement Learning. We have also seen the applications of this type of technology. This; no doubt; could be a stepping stone to the creation of machines that would out-perform the best of the human abilities, just like the case I cited about the AlphaGo that defeated the human Go champion.

Thanks for reading

References for further reading:

• deeplearning4j.org/reinforcement-learning
• technologyreview/reinforcement-learning
• Expert system/Machine-learning
• deepmind.com/meta-reinforcement-learnin

gif by @foundation