The Applications of Reinforcement Learning

By Sashrika Pandey

Image Credit: KD Nuggets

When the majority of people think of artificial intelligence (AI), the images that come to mind are the autonomous beings in Star Wars or the virtual assistants from superhero movies. While we may not yet live in a future where AI has evolved to include semi-sentient beings, there are numerous advancements being made in AI that can greatly influence your daily life. Machine learning, for instance, involves the training of a program based on past data so that it can produce a model that can be used for later analysis.

There are a plethora of subsets of machine learning, but one that stands out from the others is reinforcement learning. Reinforcement learning algorithms are unique in their training of a model since they use an underlying process known as a reward system; as a model exhibits a series of actions in response to a certain task, it is either positively or negatively affected for its actions. This reward system ensures that, through successive iterations, the model is trained to maximize the number of rewards it receives.

Take AlphaGo, which is notable for its successes in the complex game of Go. By implementing reinforcement learning, the program instead aims to maximize the rewards it receives. However, the unit of this reward may vary per scenario; as stated by Martin Heller of InfoWorld, “AlphaGo maximizes the estimated probability of an eventual win to determine its next move. It doesn’t care whether it wins by one stone or 50 stones.” While the “reward” may sometimes be numerical, it can also be the likelihood of a favorable outcome, as illustrated here. 

The underlying theory of reinforcement learning, however, emphasizes that the model is attempting to maximize the reward. Through the manipulation of several parameters and designing a model that responds to certain behaviors by increasing or decreasing a reward by a factor, researchers can ensure that the model mimics the characteristics that they are aiming for.

On a larger scale, one may take a look at Amazon, which is revolutionizing its delivery system through the use of drones that utilize reinforcement learning. According to Ron Schmelzer, Amazon “used machine learning to iterate and simulate over 50,000 configurations of drone design before choosing the optimal approach.” This particular use of reinforcement learning illustrates the efficiency through which designs and models can be adapted to a situation. Rather than a manual or basic algorithmic approach, Amazon’s use of a reward system in place of traditional methods emphasizes the possible future of reinforcement learning.

One of the most fascinating aspects of reinforcement learning is that there is a multitude of possibilities to explore and queries to test. While there are a plethora of uses for reinforcement learning currently, there are sure to be more in the future. So while we may not yet live in a society where communicating with robots is as simple as talking to other humans, we are definitely on our way to making significant advances in machine learning and artificial intelligence as a whole.

Sources:

Schmelzer, Ron. “Amazon Dives Deep into Reinforcement Learning.” Forbes. 14 Jun 2019. https://www.forbes.com/sites/cognitiveworld/2019/06/14/amazon-dives-deep-into-reinforcement-learning/#5b156f2d5fe0

Heller, Martin. “Reinforcement learning explained.” Infoworld. 2019. https://www.infoworld.com/article/3400876/reinforcement-learning-explained.html


Osiński, Błażej and Budek, Konrad. “What is reinforcement learning? The complete guide.” deepsense.ai. 5 Jul 2018. https://deepsense.ai/what-is-reinforcement-learning-the-complete-guide/