You may have heard of the term reinforcement learning (RL) in different contexts such as gaming and robotics. However, many people are not entirely clear on what RL is and the types of challenges for which it presents a great solution.

Let’s start with understanding what reinforcement learning is using a simple non-technical and real-world example: training a dog to learn new behaviours.

By presenting the dog with a treat at around the same time you create a sound (using a whistle or a clicker), gradually the dog will become conditioned to learn that the sound indicates a treat is about to be served. This is essentially “reinforcing” your dog to learn an association between a stimulus, response and reward. Over time, the dog will become accustomed to this sound and responds every time it hears it. With this technique, you can train your dog to learn certain behaviours and be rewarded for them.

Now let’s apply this “treat-and-behave” model to the context of reinforcement learning. In this scenario, the dog becomes the agent, the treat becomes the reward system, and good behavior is the ‘optimal solution’ or outcome we are seeking. The term “agent” and “reward system” might be strange to you, but we will get to them later.

What is Reinforcement Learning (RL)?

Reinforcement Learning is a branch of machine learning based on rewarding desired behaviors and punishing undesired ones. Referring to our example, one would reward the dog’s good behavior by giving him/her a treat and ‘punishing’ undesired behavior by not offering a reward. In this example, the dog is the agent, an entity who makes decisions about how to act according to rewards and punishments that are received. To put this in another way, both the start and end states are known, but there may be multiple ways for the agent to act to achieve its goal. Reinforcement learning algorithms try to find the most optimal route from this starting state to some desired outcome. To do so, a RL agent must be able to perceive and interpret its environment, take actions and learn through trial and error. By identifying which actions progress or prevent it from achieving its goal through the reinforcement of the reward system, the agent can more frequently engage in those actions that advance it towards the desired outcome.

The key to this machine learning method is the appropriate rewarding system. This method would assign positive values to the desired actions to encourage the agent and negative values to undesired ones. This technique would force the agent to seek long-term and maximum overall reward to achieve the optimal solution.

To establish the rewarding system, a q-table is often used, which lists out all the possible actions an agent could take and assign a reward to each move. Take the game flappy bird as an example. To list out all the possible moves the agent could take for this game is simple: click the mouse or don’t click. Clicking makes the bird fly up, not clicking lets the bird drop. Feeding the agent about the height of each pillar then would successfully establish the rewarding system, enabling a RL system to quickly learn how to play this game.

A breakthrough in RL

One of the most significant accomplishments of RL is AlphaGo by DeepMind. Go is a board game originated in China over 3000 years ago, which requires multiple layers of strategies. Two players, using either white or black stones, take