The Bright Future of Reinforcement Learning

Camron Godbout
3 min readMar 2, 2018


The future is bright

One of the most promising applications of deep learning is creating agents capable of making smart decisions.

The fast changing landscape of machine learning and deep learning has spread to many different applications. Deep learning has currently solved a wide range of problems, including an app that finds artwork that looks like your face, self-driving cars, and even chat bots.

One of the most interesting applications of deep learning that is making great headway is Reinforcement Learning (RL). RL combines the functional approximation of Deep Learning with the optimization of decision-based systems to bring the next wave of technology-fueled innovation.

What is Reinforcement Learning?

In traditional supervised machine learning you apply a statistical model to a dataset that includes variables that output a prediction on a given label. In reinforcement learning there is no label. But unlike unsupervised learning, where there is no label and the algorithm finds the pattern in the data, we aren’t looking for patterns. We are trying to maximize a reward function.

Reinforcement Learning loop

In reinforcement learning, we have an agent in an environment that receives observations from this environment. This agent then takes actions in the environment that returns a reward based on the action taken. The reinforcement loop is depicted in the image above.

An intuitive example to illustrate reinforcement learning is a lab rat in an experiment. We can imagine a lab rat sitting in a nice cozy box that has a string hanging from the ceiling. In this scenario the lab rat is the Agent and the environment is the cozy box. If the lab rat pulls the string, cheese falls from the ceiling and the lab rat is happy. Since the lab rat is getting rewarded every time it pulls the string it will learn to pull the string a ton of times. This is the goal of reinforcement learning — to make the agent perform a particular action to maximize it’s reward.

Apart from the lab rat example, where the desired action was trivial, we can think of applications where we want to incentivize an agent to do useful things like correctly driving a vehicle, winning games of chess, or beating humans in video games.

Why is the future bright?

Deep learning has proven its worth in approximating real-world environments. It can help solve many problems that, to date, have remained challenging for machines to tackle. With the combination of deep learning and RL, we’re much closer to solving these problems.

Future applications of reinforcement learning include many of the following tasks:

  • intelligently trade stocks and options (what we’re doing at Apteo)
  • advanced self-driving cars
  • smart traffic signals
  • autonomous robots
  • fully automated factory
  • smart prosthetic limbs

The fascinating thing about reinforcement learning is that you can take a dataset that doesn’t have any labels and make decisions to maximize your reward. This allows for unparalleled performance in environments where a certain action is desired but the means of getting there aren’t clear.


Despite all the recent advances that have been made in the field there’s still a long way to go. To see the same level of success in RL as we have in deep learning, we need the efforts of committed researchers, engineers, and data scientists. We’re still several breakthroughs away from using RL in practice, but with smart people continuing to research its capabilities, we’re making progress every day.

Finally, if any of what we’ve said has been interesting to you, you may be interested in the deep learning and RL work we’re doing at Apteo. Feel free to reach out to to learn more.