Thriving Traction | Reinforcement Learning Applied 2021-02-09T14:42:56+00:00

Thriving Traction | Reinforcement Learning Applied

Introduction

Following supervised and unsupervised learning, reinforcement learning (RL) has brought in a whole new angle for machine learning. It makes the process of learning from mistakes (like humans learn) available to machines. It enables an agent in an interactive environment to learn by trial and error, leveraging its own actions and experiences to maximize rewards. It has huge potential in data-rich sectors like robotics and games.

One of the important milestones for reinforcement learning has been its application in Google DeepMind’s AlphaGo program, which in 2016 became the first computer program to win over a human Go grandmaster.

Trained to “mimic” human players, AlphaGo competed against itself for thousands of games, improving progressively using deep reinforcement learning (DRL).

Application Across Sectors

As we move towards AGI (Artificial General Intelligence), RL has an important role to play. Let’s briefly explore some of its emerging and adapted applications:

• Deep Learning in the Oil and Gas Industry
Leading oil and gas company Royal Dutch Shell is already focusing on using deep reinforcement learning in gas extraction; it brings costs down while simultaneously improving each step along the way. The technology also studies mechanical data from drilling that is relevant to the subsurface and assists the human operator in making better decisions, thus reducing damage to expensive machinery.[1]

• Energy Optimization
Several applications, such as smart grids and data center cooling applications, have used reinforcement learning for their energy optimization. It has outperformed traditional advanced control systems and is being used to create mathematical representations of complex thermodynamic equations.[2]

• Robotics
Reinforcement learning is also being extensively used and tested in robotics. Earlier robots were incapable of using image sensing, but now the use of image data is increasingly common. DRL techniques are being used to teach robots to interact with objects, e.g. pick up, move, and drop them. However, the industry still faces challenges in moving forward.

• Healthcare
Healthcare has been quick to adopt new technologies. Most of the AI systems used in the sector focus mainly on patients’ current state and the effects of prescribed treatments when they visit the doctor’s office. DRL, on the other hand, considers both the short-term and long-term effects of medicines. However, nothing fruitful is achieved without obstacles; RL’s basic technique of trial-and-error is a hindrance when learning from real-life patients. The alternative is to learn from existing data, i.e. off-policy learning.[3]

Other areas where reinforcement learning has tremendous applicability include recommendation systems, gaming, and navigation and flight control.
Still, reinforcement learning has its own obstacles. We have come a long way in the last decade, but there are questions that must be addressed. Most DRL systems take millions of simulated steps to learn a chore, as they start from a blank slate. However, humans never start learning with a so-called blank slate; we use related experience. When a person tries to ride a motorcycle, they consider their prior experience with a bicycle. It is no small feat to try to advance RL systems by focusing on reducing simulated steps.

References

  1. Applications of DRL
  2. RL Changing the World
  3. Reinforcement Learning in Healthcare
Authored by Manshi Poonia, Data Scientist, Products & Innovation Labs, at Absolutdata